WorksPhysics and Astronomy
Finite-volume and thermodynamic-limit lower bounds for the coherently driven Bose–Hubbard model via the Lieb–Yamazaki method
Abstract
We derive rigorous lower bounds on the ground-state energy of the coherently driven Bose–Hubbard model using a Lieb–Yamazaki positive-operator decomposition. The resulting quadratic bounding Hamiltonian is diagonalized by displacement and Bogoliubov transformations, yielding explicit stability conditions and finite-volume bounds in arbitrary dimension. We establish the thermodynamic-limit energy density and characterize the excitation spectrum of the bounding Hamiltonian. We also derive its Gaussian-state observables, including filling, number fluctuations, and second-order correlations, and identify regimes of antibunching and sub-Poissonian statistics. Numerical calculations assess the tightness of the bounds across representative parameter regimes.
Introduction
The Bose–Hubbard model provides a standard setting for the competition between tunneling, local interactions, and particle-number fluctuations in lattice boson systems[1, 2, 3]. Its coherently driven form is relevant to nonlinear bosonic arrays and to protocols in which an external field controls the amplitude and phase of the lattice modes. The drive breaks particle-number conservation and introduces a linear source term, enriching the ground-state structure while removing several simplifications available in the undriven model. These features make the model pertinent to strongly correlated bosons and cold-atom physics, as well as to driven photonic lattices and quantum fluids of light[4, 5, 6, 7]. Experiments on open photonic systems have observed dissipative critical behavior and stabilized Mott states[8, 9], while driven-dissipative theory has addressed critical slowing down, excitation and correlation structure, and chiral or chaotic dynamics[10, 11, 12, 13]. Open photonic realizations motivate the coherent source term, but the present work addresses the closed Hamiltonian sector only; dissipation and steady-state Liouvillian spectra are outside its scope.
Variational trial states readily furnish upper bounds on the ground-state energy, but rigorous lower bounds are considerably less accessible in the presence of both a quartic on-site interaction and a coherent source. A useful lower estimate must control the unbounded local occupation, remain stable under the thermodynamic limit, and retain explicit dependence on the detuning, hopping, and drive. It is also desirable to determine what the bounding construction implies for observables and excitations, rather than for the energy alone. Rigorous Bogoliubov analyses and energy estimates for interacting Bose gases provide complementary control of quadratic approximations and their remainders[14, 15, 16]. Elementary one-site and bond-cluster estimates also follow from operator inequalities and Anderson’s cluster lower-bound construction [17]. We therefore compare the Lieb–Yamazaki result directly with these important baselines rather than claiming uniqueness of the bound.
Herein, the Lieb–Yamazaki positive-operator decomposition is applied[18] to the driven Bose–Hubbard Hamiltonian. Upon the subtraction of a sum of positive operators, a quadratic bounding Hamiltonian is left, and is subsequently reduced by a uniform displacement and a Bogoliubov transformation[19, 20]. An admissible family of decomposition parameters is thereby delimited by the resulting stability condition. By the min–max principle, the quasiparticle gap of the bounding Hamiltonian is then converted into a certified lower threshold for the first excited energy of the full model.
From the construction there is obtained the rigorous lower bound of Theorem II.3 in every dimension and its explicit thermodynamic-limit energy density. Proposition III.2 identifies the continuous excitation band of the bounding Hamiltonian, whilst from the Bogoliubov vacuum there follow formulas that are exact within the Gaussian ground state of the bounding Hamiltonian for the order parameter, filling, energy densities, particle-number fluctuations, anomalous averages, and a drive-induced antibunching criterion. Finally, the certified Temple analysis of § 7 clarifies the finite-volume condition under which this Gaussian ansatz can support a Temple correction. The lower-bound property, per se, does not depend upon that correction. Its extensive energy excess makes that condition increasingly restrictive with system size, independently of the quality of the lower energy-density bound.
The Model and the Lieb–Yamazaki Procedure
Model Hamiltonian
We consider an array of coupled single-mode microcavities containing nonlinear media and driven coherently by external radiation. In a frame rotating at the driving frequency, the driven Bose–Hubbard Hamiltonian is
Here, the bosonic operators obey . The parameter is the on-site Kerr nonlinearity, is the nearest-neighbor hopping amplitude, is the coherent driving amplitude, and is the detuning between the drive and the bare cavity mode. The lattice contains sites and is endowed with periodic boundary conditions. The notation denotes nearest-neighbor pairs, counted once. Throughout we assume a repulsive interaction . A gauge transformation makes the hopping nonnegative, so we take ; a global phase rotation likewise permits real .
Rescaling to Dimensionless Variables
For , we measure all energies in units of and introduce the dimensionless quantities
The stationary Schrödinger equation then takes the form
In these variables, the on-site interaction has unit coefficient, which makes the algebraic structure used below explicit.
Lieb–Yamazaki Decomposition
Following the positive-operator strategy of Lieb and Yamazaki18, we introduce at every lattice site
where and are real variational parameters. The operator
is positive semidefinite. The remainder in the exact decomposition below is
Proposition II.1 (Positive-Operator Decomposition). Let for real parameters and . Then the dimensionless Hamiltonian admits the exact decomposition
where and is given explicitly by Eq. (6). Consequently, as an operator inequality.
Proof of Proposition II.1. On the common invariant finite-particle core, direct expansion of , using , gives Eq. (7) term by term. Moreover, for every , so that summation over gives ; hence the asserted inequality holds there as a quadratic-form inequality and extends to the corresponding closed forms21. This is the positive-operator step of the Lieb–Yamazaki construction18.
Uniform c-number Shift
The linear terms are eliminated by the real, spatially uniform displacement ,
While the original polaron application of Lieb and Yamazaki employs a momentum-dependent coupling , we adopt the uniform ansatz . This choice preserves the diagonal momentum-space structure of the Bogoliubov transformation. It permits an analytic treatment of the thermodynamic limit and closed-form expressions for the energy density and excitation spectrum. The same framework also permits momentum-dependent couplings.
With the dimensionless detuning of Eq. (2), define
Cancellation of the linear terms fixes the displacement to
Substitution into Eq. (6) gives
The last term is the complete c-number contribution, including the term generated by the Lieb–Yamazaki decomposition.
Fourier Transformation
Translation invariance permits the discrete Fourier representation
where and all momentum indices are understood modulo . The quadratic lower-bound Hamiltonian becomes
where
A positive normal coefficient requires for every momentum . The stronger stability condition needed by the Bogoliubov transformation is obtained below.
Bogoliubov Transformation
We introduce new canonical bosonic modes through
where is real and . Choosing to cancel the anomalous terms diagonalizes Eq. (13) and yields
The physical branch, continuously connected to as , is
Lemma II.2 (Stability Condition for Bogoliubov Transformation). For , the physical branch of the Bogoliubov transformation defined by Eq. (15) is canonical and diagonalizes a positive quadratic mode Hamiltonian if and only if
For , this condition is equivalent to the constraint on given by Eq. (23).
Proof of Lemma II.2. The Bogoliubov transformation is canonical if and only if its coefficients satisfy and the transformation preserves the commutation relations. Reality of the square root alone permits either or . The latter branch has a negative coefficient of the number operator after diagonalization and is therefore unbounded below. Positivity selects , for which the physical root in Eq. (17) also obeys . For , the minimum of occurs at , which yields the stated constraint on . When , one instead takes directly; the positive-mode condition is . This is the standard bosonic Bogoliubov criterion [19, 20].
Extension to Dimensions
The construction above does not depend on the dimension of the lattice. Let and let be the periodic hypercubic lattice of linear size , with sites labeled by . We consider
which reduces to Eq. (1) for . The hopping amplitude per spatial direction is thus . With this normalization, the hopping energy of the zero-momentum mode is independent of . The dimensionless quantities are defined as in Eq. (2).
The steps of the one-dimensional derivation carry over as follows.
The operator of Eq. (4) acts on a single site, and the hopping term is not affected by the decomposition. Proposition II.1 therefore holds verbatim, with sums over and with the constant in .
Every site belongs to nearest-neighbor pairs, each carrying the coefficient . The uniform displacement of Eq. (8) thus produces the same linear terms per site as in one dimension, and Eq. (10) for is unchanged. The c-number contribution becomes .
The Fourier modes are labeled by ,
and Eqs. (13)–(17) hold with replaced by , with replaced by (modulo in each component), with the constant replaced by , and with
Here is the free-particle dispersion, with for every .
The excitation energies and the Bogoliubov coefficients are
so that ; for both sides vanish.
For , the minimum of is attained at , where all cosines equal one, and equals . The stability condition of Lemma II.2 is therefore equivalent to the same constraint on in every dimension.
Ground-State Energy Lower Bound
Theorem II.3 (Rigorous Lower Bound in Dimensions). Let , let be the driven Bose–Hubbard Hamiltonian on defined by Eq. (19), and let be the corresponding dimensionless Hamiltonian, with . For any real parameters and satisfying
the ground-state energy of satisfies the rigorous lower bound
where , the sum runs over the modes , and is given by Eqs. (21) and (22). Equivalently,
The constraint in Eq. (23) does not depend on . For , Eq. (24) reads , with given by Eq. (17).
Proof of Theorem II.3. Proposition II.1 gives in the quadratic-form sense. The variational (min–max) principle[22] therefore implies . The Hamiltonian is quadratic in the bosonic operators and is diagonalized by the Bogoliubov transformation in Eq. (15), which is well-defined by Lemma II.2. The vacuum of the modes minimizes the diagonalized , and its lowest eigenvalue is given by Eq. (24); Eq. (25) follows from Eq. (22). The dimension enters this argument only through the number of sites, , and through the sum over modes: the Lieb–Yamazaki decomposition is local, acting on one site at a time, and the remaining steps were adapted to in § 2 G. The proof is thus the same for every . □
Summary
It follows that, in every dimension , there is a variational family of rigorous lower bounds for the ground-state energy, parameterized by and subject to the stability constraint. In the following section, we take the thermodynamic limit to extract the energy density and excitation spectrum.
THE THERMODYNAMIC LIMIT
Strategy: From Finite Lattice to Thermodynamic Limit
Theorem II.3 provides a rigorous lower bound for every finite periodic lattice of sites, in any dimension . Its derivation uses the positive-operator inequality of Proposition II.1 and the stability criterion of Lemma II.2. We now prove the existence of the corresponding energy density
It is essential that passage to the thermodynamic limit preserve the finite-volume lower-bound property. Let denote the momentum variable of the -dimensional Brillouin zone. If is continuous and -periodic in each component, the standard convergence theorem for Riemann sums on the grids gives
see, e.g., Ref.23.
Energy Density in the Thermodynamic Limit
Dividing the finite-volume bound in Eq. (24) by yields
Introduce the continuous functions
and
The function is the root of with , and . For , the expression is understood by its continuous limit as , which is zero. By Eqs. (21) and (22), , , and, for , .
Corollary III.1 (Energy Density in the Thermodynamic Limit). Let and . For every admissible pair satisfying the stability constraint in Eq. (23), the thermodynamic limit of the finite-volume lower-bound density exists and is
Equivalently,
Without assuming existence of the exact thermodynamic limit, this expression obeys
Proof of Corollary III.1. For admissible parameters, for every , and hence and is continuous and -periodic in each component. The Riemann-sum limit in Eq. (27) applies with . The remaining terms in Eq. (28) are independent of , which proves Eq. (31). Equation (32) follows from , because the cosines integrate to zero and the Brillouin-zone average of is therefore . Moreover, Theorem II.3 gives for every . Taking the lower limit proves Eq. (33); if the exact energy density exists, this reduces to the usual thermodynamic-limit inequality.
Continuous Bogoliubov Excitation Spectrum
Proposition III.2 (Continuous Excitation Spectrum). The diagonal coefficient in Eq. (16) is . In the thermodynamic limit these Bogoliubov excitation energies sample the continuous dispersion
The stability condition becomes for every , which for is equivalent to the constraint on in Eq. (23), independently of .
Proof of Proposition III.2. Substitution of into the coefficient of in Eq. (16) gives above, since solves . The momentum grids become dense in the Brillouin zone; continuity of in the stable region then gives the claimed continuum band. For , the minimum of occurs at and equals in every dimension, so uniform stability is precisely the constraint in Eq. (23).
Analysis of the Energy Density Integral
For every admissible pair , the strict inequality ensures that the radicand in Eq. (30) is positive throughout the compact Brillouin zone. Consequently, is smooth in , and the integral in Eq. (31) is well-defined. It may be evaluated numerically for general parameters and analytically in special limits.
At the boundary of the admissible region, there may exist a critical momentum for which . The square root then becomes nonanalytic. Such behavior marks the loss of Bogoliubov stability of the bounding Hamiltonian; it is not evidence of a phase transition of the full driven model. The rigorous bound derived above applies in the open stable region; any extension to its boundary requires a separate limiting argument.
Summary
We have thus rigorously established the thermodynamic limit of the Lieb–Yamazaki lower bound in every dimension . The lower-bound density is given by a closed-form integral expression, and the excitation spectrum is described by the continuous dispersion relation . In the following section, we derive macroscopic observables from this framework.
Excitation Spectrum and Observables
Order Parameter, Filling Fraction, and Energy Densities
Throughout this subsection, we take ; the extension to dimensions is given in Remark IV.8. The pair is admissible in the sense of Eq. (23), and denotes the normalized vacuum of the Bogoliubov modes of Eq. (15),
By the proof of Theorem II.3, is the ground state of the bounding Hamiltonian ; it depends on and . All expectation values below are taken in this state, , and all energies are measured in the dimensionless units of Eq. (2). We write as in Proposition III.2.
Lemma IV.1 (Two-Point Functions). In the state , and
with momentum indices understood modulo . Consequently, the local moments of the displaced operators are independent of the site and equal
while $\langle\tilde a_k\rangle=0.
Proof of Lemma IV.1. (15) expresses as a linear combination of and , so that by Eq. (35). Using and , we have
Since in Eq. (17) solves , it follows that , which converts Eq. (38) into Eq. (36). For one has and , and both sides of Eq. (36) vanish. Inserting the Fourier representation in Eq. (12) yields Eq. (37); the phases cancel in both sums, which proves the independence of .
Proposition IV.2 (Order Parameter). In the state , the expectation value of the annihilation operator is the same at every site and equals the displacement,
where as in Eq. (9).
Proof of Proposition IV.2. By Eq. (8), , and by Lemma IV.1. The value of is given by Eq. (10).
Proposition IV.3 (Filling Fraction). In the state , the mean number of particles per site is
with given by Eq. (39).
Proof of Proposition IV.3. Since is real, Eq. (8) gives . The linear terms have vanishing expectation, and the last term equals by Lemma IV.1.
The first term in Eq. (40) is the coherent contribution of the displacement, while the second is the depletion generated by the anomalous terms of ; the latter vanishes for .
Proposition IV.4 (Kinetic Energy Density). Let
be the hopping part of the dimensionless Hamiltonian . In the state ,
where is the free-particle dispersion and .
Proof of Proposition IV.4. Let denote the Fourier modes of the undisplaced operators , defined as in Eq. (12). Then , and Eq. (8) gives . By Lemma IV.1, . Summation over and division by prove the claim.
Proposition IV.5 (On-Site Energy Density). Let
be the on-site part of , so that . In the state ,
where and are the local moments in Eq. (37). This expression is exact within the Gaussian ground state of the bounding Hamiltonian .
Proof of Proposition IV.5. The detuning and driving terms follow from Propositions IV.2 and IV.3. For the interaction term, Eq. (8) gives
The state is the vacuum of operators that are linear in and , hence it is quasi-free and Wick’s theorem applies1. Expectation values of odd monomials vanish, and because is real. The remaining even terms of Eq. (45) contribute .
Remark IV.6 (Strong-Condensate Limit). If the displacement dominates the fluctuations, , the terms in Eq. (44) may be neglected. To first order in the fluctuation moments one then has the approximation
In contrast with Eq. (44), this is not an identity; it is valid only in the strong-condensate limit.
Corollary IV.7 (Observables in the Thermodynamic Limit). For every admissible pair , the limits of the quantities above exist. With and as in Eq. (29) for ,
and the limit of is given by Eq. (44) with and replaced by and . The order parameter is independent of .
Proof of Corollary IV.7. In the stable region, for every , so the integrands are continuous and -periodic. The finite-volume expressions are their Riemann sums on the grid , and the convergence argument used for Eq. (27) applies.
Remark IV.8 (Extension to Dimensions). On the lattice of § 2 G, the results of this subsection hold with the following replacements:
the mode averages become , with and replaced by and of Eqs. (21) and (22), and with in Eq. (36) replaced by ;
the free-particle dispersion becomes of Eq. (21), the hopping operator in Eq. (41) carries the coefficient , and becomes ;
the Brillouin-zone integrals and the cosine in Eq. (48) become
The order parameter and the expressions of Propositions IV.3 and IV.5 and of Remark IV.6 in terms of , , and do not change, because they are local and does not depend on . All integrands remain continuous, since in the stable region.
Remark IV.9 (Variational Upper Bound). Since and is a normalized trial state, Propositions IV.4 and IV.5 also give a variational upper bound on the ground-state energy. Combined with Theorem II.3 for ,
and by Proposition II.1 the difference between the two outer expressions is . In dimensions the same statement holds with and in place of and .
Particle-Number Variance and Correlations {#sec:B}
We now turn from one-body expectation values to local two-body quantities in the state : the variance of the on-site particle number, the second-order correlation function, and the anomalous averages. Control of these fluctuations is a prerequisite for assessing the quality of the lower bound in § 7. As in § 5 A, we take ; since every result below is expressed through , , and , Remark IV.8 applies without change. Because is quasi-free, all formulas of this subsection follow from the two-point functions of Lemma IV.1 by Wick’s theorem and are exact, with the exception of Remark IV.14.
We write and recall from the proof of Proposition IV.5 that
Proof of Lemma IV.10. For and admissible parameters, and , which gives and term by term. Moreover,
and , so every term of the sum is negative. For , and both moments vanish.
Proposition IV.11 (Particle-Number Variance). In the state , the variance of the on-site particle number is
where is the filling fraction of Eq. (40). This expression is exact within , the Gaussian ground state of .
Proof of Proposition IV.11. The commutation relation gives the operator identity , hence . Subtracting from Eq. (50) yields Eq. (53).
Two limiting cases illustrate Eq. (53). For , and , the Poissonian value of a coherent state. For , , the super-Poissonian value of a squeezed vacuum. In general, Lemma IV.10 shows that the term is negative for , so that the number fluctuations are sub-Poissonian, , if and only if
Proposition IV.12 (Second-Order Correlation Function). In the state with , the local second-order correlation function is
This expression is exact within , the Gaussian ground state of , and .
Proof of Proposition IV.12. The first equality is Eq. (50) divided by , with given by Proposition IV.3; the second follows by subtracting in the numerator. The last statement is a consequence of Proposition IV.11.
Values correspond to bunching, to Poissonian statistics, and to antibunching. For and , Eq. (55) gives , as for a coherent state. For and , it gives , which exceeds 3 because by Lemma IV.10; this is the strong bunching of a squeezed vacuum. Since for , a sufficiently large displacement reverses the sign of : the state is antibunched precisely when Eq. (54) holds. As is proportional to by Eq. (39), this antibunching is induced by the coherent drive. The use of as a blockade and nonclassicality diagnostic parallels recent studies of dynamical, multi-emitter, and squeezing-induced bosonic blockade[24, 25, 26].
Proposition IV.13 (Anomalous Averages). In the state ,
Proof of Proposition IV.13. By Eq. (8), . The linear term has vanishing expectation and by Lemma IV.1. The two averages coincide because and are real.
The difference is the anomalous average of the fluctuations. It originates from the pairing of the modes and in Eq. (36); it is nonzero for and vanishes for .
Remark IV.14 (Strong-Condensate Limit of the Fluctuations). If , the terms in Eqs. (53) and (55) may be neglected. To first order in the fluctuation moments one then has the approximations
In contrast with Propositions IV.11 and IV.12, these are not identities; they are valid only in the strong-condensate limit. By Lemma IV.10, in this limit approaches unity from below whenever .
Corollary IV.15 (Fluctuations in the Thermodynamic Limit). For every admissible pair , the limits of , (for , and exist. They are given by Eqs. (53), (55), and (56) with and replaced by the integrals and of Eq. (47). The inequalities of Lemma IV.10 hold for and as well.
Proof of Corollary IV.15. The three quantities are continuous functions of at fixed , and , by Corollary IV.7. The integrands of , , and have the signs established pointwise in the proof of Lemma IV.10.
Boundary Slope of the Bounding Hamiltonian
We finally examine the long-wavelength behavior of the continuous band of Proposition III.2, with and as in Eqs. (29) and (34). In contrast with § 5 A and IV B, this subsection is formulated directly for arbitrary . We assume ; for the band is flat. Because is a function of the continuous momentum , every statement below already refers to the thermodynamic limit, and no separate limiting corollary is needed. We use
Proposition IV.16 (Boundary Slope). Let and .
In the stable region , the excitation spectrum has a gap, attained at :
As ,
so that the dispersion is quadratic and the ratio diverges as : there is no finite linear boundary slope.
On the boundary of the stable region, with , the function of Eq. (34) remains well defined and
The gap is closed and the dispersion is linear at small momenta,
Proof of Proposition IV.16. By Eq. (29), with , and . Hence
In the stable region , so the right-hand side of the first equality is increasing in and is minimal at , which gives Eq. (59). Since , the square root is smooth near , and its expansion gives Eq. (60). On the boundary, and , and the first equality in Eq. (63) reduces to Eq. (61). Then , and taking the square root gives Eq. (62).
The two regimes are connected by Eq. (63), which has the relativistic form with , so that . In the stable region the excitations of the bounding Hamiltonian are massive quasiparticles with effective mass ; the dispersion is quadratic for and approximately linear with slope above this scale. As the boundary is approached, , , and the quadratic window shrinks to the point ; the limits and therefore do not commute. The velocity depends only on the hopping amplitude , the parameter , and the dimension; it does not depend on the detuning or on the drive amplitude , because on the boundary is fixed by .
Two limitations should be kept in mind. First, and are parameters of the Lieb–Yamazaki decomposition and not of the model, so the gap, , and characterize the bounding Hamiltonian for a given choice of ; the boundary is the edge of the admissible region and not a phase boundary of the driven Bose–Hubbard model. Second, Theorem II.3 and the results of § 5 A and IV B are proved in the open stable region only. On the boundary the integrands of Eq. (47) behave as near , which is integrable for but not for , where and diverge logarithmically as the boundary is approached.
Remark IV.17 (Comparison with the Case ). For the spectrum is , the stable region is , and Eq. (60) gives the free lattice mass . On the boundary one has : the gap closes, but the dispersion remains quadratic and the boundary slope vanishes, in agreement with Eq. (62). Moreover, the displacement of Eq. (10) diverges there unless . For , in contrast, is finite on the boundary and the gapless excitations have the linear dispersion of Eq. (62). A finite linear boundary slope at the edge of the stable region is thus a qualitative effect of the pairing terms generated by .
Numerical Illustration of the Spectrum and Local Correlations
Figure 6(a) illustrates the two long-wavelength regimes of Proposition IV.16 for . The deep-stable curve uses the independently optimized variational parameters at , and ; its stability margin is approximately 0.09144 and its gap is , so the dispersion is quadratic at sufficiently small momentum. For comparison, the near-boundary curve uses and , giving and a small residual gap . Outside the correspondingly narrow quadratic window it follows the boundary law , with
Figure 6. Excitation spectrum and local correlations at . (a) The solid curve with open-circle markers denotes the gapped, quadratic long-wavelength dispersion at the optimized deep-stable point for ; the dashed curve with square markers denotes a point having stability margin , and the dotted curve is the bounding-Hamiltonian boundary asymptote for . (b) The on-site second-order correlation of the stable Gaussian family with fixed as the coherent drive is increased. The dashed line marks Poissonian statistics; values above and below it indicate bunching and antibunching, respectively.
This linear dispersion defines a boundary slope, which may be viewed as an effective sound velocity of the excitations of the bounding Hamiltonian . It is not a sound velocity established for the full driven model and does not mark a phase transition of that model. Whether the full interacting Hamiltonian has a corresponding low-energy feature must be tested independently, for example by finite-size scaling or tensor-network methods. Collective-mode spectroscopy of Bose–Hubbard systems likewise distinguishes low-energy phase modes from massive amplitude response[27, 28].
Figure 6(b) holds the same stable near-boundary pair fixed and varies only the coherent drive. This isolates the displacement-induced change in the Gaussian trial state. The moments and are evaluated from Eq. (47), the filling and number variance from Eqs. (48) and (53), and from Eq. (55). The correlation starts in the bunched regime, reaches its squeezed-vacuum maximum near , and crosses the Poissonian value between and 1.50. At stronger drive it remains below unity (for example, its minimum is approximately 0.9246 near ), demonstrating drive-induced antibunching and sub-Poissonian fluctuations. Thus exposes the nonclassical number statistics of this coherently displaced auxiliary Gaussian state; it is not, by itself, a prediction for the ground state of the full interacting Hamiltonian.
Numerical Optimization and Tightness of Bounds
Let be the normalized ground state of the bounding Hamiltonian , and write the dimensionless Hamiltonian as . Its expectation in this state is a variational upper bound on the ground-state energy .
Lemma V.1 (Certified Gap for the Full Hamiltonian). Let be the full dimensionless Hamiltonian and the bounding Hamiltonian of Eq. (16). In finite volume and in the stable region, the first excited-state energy of satisfies
Proof of Lemma V.1. By Proposition II.1, in the quadratic-form sense. The min–max principle2 therefore gives for every . The diagonal form of has spectrum , where . Hence by Proposition IV.16, proving the claim. □
We next give an exact and directly implementable expression for the full variance. For self-adjoint operators and , define
With , the exact variance decomposition is
In a translation-invariant state on the periodic lattice, the last term is
Consequently it cannot in general be replaced by : that replacement would require all inter-site covariances to vanish.
For completeness, define and the translation-invariant contractions
which are the inverse discrete Fourier transforms of, respectively, and . Thus and . If is any even word in the , put
where the four ordered contractions follow from and . Expanding
as , define the completely explicit bilinear Wick sum
All words in Eq. (73) have length at most eight.
Proposition V.2 (Energy Variance in the Bogoliubov Vacuum). In the state , the exact full-Hamiltonian variance is
Equivalently, with ,
Equations (70)–(74) are finite lattice sums in , , , , and give the requested exact expressions for , , and .
Proof of Proposition V.2. The first identity is the algebraic variance decomposition. The state is quasi-free, so Wick’s theorem evaluates every term exactly as Eq. (73). Finally and ; consequently constants and all mixed covariances with drop out, giving Eq. (75).
Theorem V.3 (Temple’s Inequality with Certified Gap). Let be the variational upper bound of Remark IV.9. If , then
Together with Theorem II.3, this gives
Proof of Theorem V.3. Lemma V.1 supplies the lower spectral threshold . The stated condition implies , so Temple’s inequality gives Eq. (76). Taking the stronger of this result and Theorem II.3, and using the variational principle for the upper bound, proves Eq. (77).
Elementary Site and Bond Baselines
The Lieb–Yamazaki estimate should be judged against two elementary rigorous lower bounds. Define the one-site ground energy
and the two-site bond Hamiltonian
Then the exact energy density satisfies
Indeed, follows from ; summing bonds gives the site estimate. For the second estimate, and there are bonds, so the Anderson cluster argument [17] applies.
Table 1 reports a cutoff-converged numerical comparison in one dimension. The bond bound is substantially sharper than the Lieb–Yamazaki bound at every listed point, while the site bound can also be competitive. Thus the value of the present construction lies primarily in its analytic parameter dependence and its explicitly solvable auxiliary Gaussian state, not in uniformly optimal energy bounds.
| LY | site | bond | ED | |
Table 1. Lower-bound energy densities in ; larger values are sharper. The last column is the converged truncated-Hamiltonian reference.
Optimization of the Variational Parameters
We now optimize the thermodynamic lower-bound energy density of Corollary III.1 over the admissible decomposition parameters. For fixed model parameters , the formal optimum is the supremum over the open stable set,
For numerical work we keep a distance from the boundary and define
and denote an optimizing pair by . The parameters and belong to the Lieb–Yamazaki decomposition; this optimization is over those variational parameters, not a limit such as . We use in the numerical calculations. At regular boundary
points the energy density has a finite continuous limit, so this regulator is not part of the formal supremum; it stabilizes the numerical quadrature. In one dimension, however, individual Bogoliubov moments and the same-pair upper estimate can diverge or depend logarithmically on near a gapless optimizer. We distinguish two upper-bound energy-density protocols. The upper-bound energy density evaluated at the lower-bound optimizer is
whereas independent upper-bound energy-density optimization gives
For brevity, figures label as . The associated is explicitly regulator dependent, whereas is the primary upper estimate used in the regime maps. All numerical multistart searches, including the independent minimization in Eq. (84), are restricted to ; their reported best points are not proofs of global extrema. Only the same-pair gap obeys the decomposition identity
where the equality follows from the positive-operator decomposition and the fact that is the ground state of . For the independently optimized bracket one has only by the variational and lower-bound principles.
Figure 1 compares the optimized thermodynamic lower-bound energy density with the upper-bound energy density throughout at fixed for the reference hopping , weak hopping , and strong hopping . Numerically converged eigenvalues of the -truncated periodic chain are included in every panel. We retain outside the mixed strong regime; every marker with and was recomputed adaptively at until consecutive energy densities differed by less than . These points are exact for their finite-dimensional truncated Hamiltonians and serve as numerical reference values, rather than exact thermodynamic energy densities. Exact diagonalization is restricted by this rapid Hilbert-space growth; matrix-product-state and DMRG methods provide the standard scalable alternative in one dimension [29, 30, 31, 32, 33]. In panel (a), at , the optimization gives , , , and , so the relative same-pair gap is about . This number is regulator dependent near the stability boundary. Unless independent minimization is stated explicitly, all gaps discussed for Figure 1–Figure 3 use this same-pair protocol. Comparing the panels shows that weak hopping produces the narrowest relative envelope at the upper end of the scan: at the relative gaps are approximately , , and for , , and , respectively. Thus all three envelopes tighten with increasing drive, although the approach is slowest in the strong-hopping case. The color map in Figure 5 shows the lower-bound landscape at and marks its numerically found maximum in the stable region.
Figure 1. Optimized thermodynamic lower-bound energy density and variational upper-bound energy density $e_u^{\mathrm{at\mathchar-lower}}$ evaluated at the same lower-bound optimizer versus drive amplitude over $0 \le f \le 5$ for (a) $j = 1.0$, (b) $j = 0.2$, and (c) $j = 5.0$, at fixed $\nu_0 = 1$. Small black dots are converged eigenvalues of the $n_{\max} = 10$ truncated periodic chain with $M = 4$ outside the mixed strong regime; black crosses for $f \ge 4$, $j \ge 4$ use adaptive $n_{\max} = 14$–$18$ with successive energy-density changes below $10^{-4}$ (see {xref}Fig. 8`). These finite-chain markers are numerical references, not TDL energy densities. The numerical step is .
Figure 3. Optimized thermodynamic lower-bound energy density and variational upper-bound energy density evaluated at the same lower-bound optimizer versus detuning over , beginning at exact resonance, for (a) , (b) , and (c) , at fixed strong drive . Small black dots are converged eigenvalues of the truncated periodic chain with for and ; black crosses for use adaptive – with successive energy-density changes below (see Figure 8). These finite-chain markers are numerical references, not TDL energy densities. The numerical step is .
Figure 5. Thermodynamic lower-bound energy-density landscape for and . The hatched region is outside the stability constraint regularized by , and the star indicates the numerical maximum. The optimization uses a nine-start L-BFGS-B search over inside the stable region.
The larger cutoff is essential in the mixed strong regime. At , increasing from 6 to 10 lowers the truncated-chain energy density from approximately to ; the adaptive calculation converges at to approximately . The converged value lies between the thermodynamic lower and upper curves. Thus the earlier points above the upper curve were a local-Hilbert-space truncation artifact. The reduced cutoff error makes placement inside the thermodynamic envelope a reliable numerical diagnostic at the plotted resolution.
Figure 2 gives the complementary scan in the hopping parameter at fixed drive. For weak drive, , the absolute gap has a shallow minimum of approximately near and then broadens substantially to about at . At strong drive, , it grows much more gently, from approximately at to at ; relative to , the corresponding endpoint gaps are about and . At every sampled point the finite-chain ED marker lies between the two thermodynamic curves.
Figure 2. Optimized thermodynamic lower-bound energy density and variational upper-bound energy density evaluated at the same lower-bound optimizer versus hopping parameter over for (a) , (b) , and (c) , at fixed . Small black dots are converged eigenvalues of the truncated periodic chain with outside the mixed strong regime; black crosses for , use adaptive – with successive energy-density changes below (see Figure 8). These finite-chain markers are numerical references, not TDL energy densities. The numerical step is .
Figure 3 starts at exact resonance, , and tracks the bounds with increasing detuning at fixed strong drive . At resonance the absolute gaps are approximately 0.228, 0.320, and 1.800 for , 1, and 5, respectively. The corresponding stability margins are approximately 0.7251, 0.3108, and , so every optimum is stable; the strong-hopping solution lies close to the regularized stability boundary. The absolute gaps increase monotonically over the plotted detuning interval, reaching approximately 0.786, 1.423, and 2.739 at . All finite-chain ED points again lie between the two thermodynamic curves throughout the scan.
Figure 4 separates finite-size effects from the thermodynamic approximation at . The finite-volume Lieb–Yamazaki lower bound uses the four discrete momenta of the ring and rigorously bounds its exact ground-state energy from below. Evaluating the finite-volume variational energy and minimizing it independently over and in the stable region supplies the corresponding optimized upper bound. The thermodynamic upper curve is minimized in the same way. The separation of each lower–upper pair is therefore a direct measure of the best bracket obtained from the Bogoliubov ansatz for that system size, independently of the thermodynamic limit. Numerically, every recomputed ED point lies inside this finite-volume bracket over ; the smallest observed distances from the lower and upper curves are approximately 0.144 and 0.179, respectively. The cutoff ED values approximate the untruncated finite-volume energy to which the rigorous bounds apply. The finite-volume and thermodynamic upper bounds differ by at most approximately 0.0100 over the scan and their curves are nearly superposed. The same is true of the lower bounds, whose maximum absolute difference is approximately 0.0120. Thus finite-size corrections to both optimized bounds are small for in this parameter range, even though the finite-volume curves are the ones that rigorously bracket the finite system.
Figure 4. Finite-size comparison at , , and . Blue solid and dashed curves are the optimized thermodynamic Lieb–Yamazaki lower-bound energy density and independently minimized variational upper-bound energy density . Red dotted and dash-dotted curves are their finite-volume counterparts. Each upper curve is independently minimized over and inside the stable region, while each lower curve is independently maximized there. Black points are converged eigenvalues of the truncated Hamiltonian for ; their changes from are below , well within the criterion of Fig. 8. These finite-chain markers are numerical references, not TDL energy densities. The finite-volume pair isolates the ansatz error at fixed size, while its displacement from the blue curves displays finite-size effects.
The four curves use four separate optimizing pairs at every drive. For example, Table II gives the values at . The lower-bound pairs maximize the corresponding Lieb–Yamazaki functional, whereas the upper-bound pairs independently minimize the variational energy; all four optimizations are restricted to the stable region.
| Bound | ||
| Thermodynamic lower | 1.540123 | 0.141461 |
| Thermodynamic upper | 1.753143 | −0.341058 |
| Finite-volume lower, | 1.539911 | 0.141042 |
| Finite-volume upper, | 1.745619 | −0.323582 |
Table II. The four independently optimized stable parameter pairs at , , and .
Numerical Assessment of the Temple Estimate
For a translation-invariant Bogoliubov vacuum, the decomposition gives
The Temple applicability condition is therefore
Unless vanishes with volume, its left-hand side is extensive while the certified one-particle gap is intensive. Hence a fixed trial family can satisfy the test only below a parameter-dependent size threshold; closeness to the stability boundary makes the condition still more restrictive because closes. At a given point, the auxiliary gap would have to exceed ; equivalently, if the present test fails, it must increase by a factor greater than while the excess is held fixed. This is a necessary threshold, not evidence that any particular extension of the ansatz will achieve it. A finite cutoff imposed directly on the Gaussian state can also spoil this identity, so numerical tests must evaluate the Gaussian moments consistently, for example through the Wick formulas above.
This scaling observation replaces a universal negative verdict based on a small cutoff scan. At a particular finite size the condition may pass or fail, and a different admissible decomposition may improve the threshold, but neither outcome by itself diagnoses the spectrum of the full model. Equation (76) remains a valid conditional finite-volume refinement; it does not yield a thermodynamic energy-density correction for a nonzero limiting .
Regime Maps and Limitations of the Ansatz
The four regime maps in Figure 7 summarize where the optimized Bogoliubov construction is quantitatively tight. Panels (a) and (c) compare the plane at detunings and , while panels (b) and (d) compare the plane over at hoppings and . The relative gap is below throughout all four sampled planes by ; the numerical contour is already reached between approximately and , depending on the remaining parameter. Consequently the especially conservative statement that the bounds are narrow for strong drive and moderate hopping is satisfied, as is the corresponding statement for and . The map in fact shows that the latter precision extends over the full sampled hopping and detuning ranges at .
Figure 7. Regime maps of the optimized thermodynamic bounds on (a) the plane at , (b) the plane at , (c) the plane at exact resonance , and (d) the plane at weak hopping . Both detuning panels cover . Color, with logarithmic color normalization, represents the relative bracket width , where is independently minimized, clipped at 50%. Hatched regions have a relative bracket exceeding 50%. Thick dashed and dotted contours denote, respectively, the stability margin and a relative gap of 10%. The haloed solid contour denotes ; the white dash-dotted lines in the detuning panels mark exact resonance. Every displayed optimum is stable, so no hatched unstable region is present.
The red dashed contour marks a stability margin of . All optimized points shown are inside the stable region, as required by the construction; the contour identifies the narrow strip in which the solution approaches its boundary. At large hopping, in particular near , the lower-bound optimizer remains close to this contour over the displayed drive interval, which helps explain the comparatively wider bracket there. Crossing the actual boundary would make the Bogoliubov frequencies imaginary and the transformation physically inadmissible.
The black contours in Figure 7 mark for the Gaussian state selected by the lower-bound optimum. In panels (a) and (c) they separate a bunched region at weak hopping and weak drive from the antibunched region generated by coherent driving; exact resonance enlarges the bunched region. At weak hopping , panel (d) similarly exhibits a pronounced bunched tongue at small drive. Extending the detuning range to negative values also reveals a bunched pocket in panel (b), centered on weak drive and negative detuning, which was absent from the previously displayed nonnegative-detuning window. These are correlation crossovers, not phase transitions.
As a reading guide, the upper portions of all four panels () are the narrow-bracket regime, below the dotted contour. The low-drive parts contain the widest brackets and all crossings of the haloed contour, while the dashed stability-margin contour identifies where the optimized auxiliary Hamiltonian approaches its gapless boundary.
The map should not be read as a conventional equilibrium phase diagram. In particular, this method does not reproduce the Mott lobes of the undriven Bose–Hubbard model. The coherent term explicitly breaks the number-conserving symmetry, while the uniform displacement builds the ansatz around a coherent state. Thus every induces some coherence and removes the sharp symmetry-based Mott transition, although smooth remnants and crossovers may remain.
The Gaussian Bogoliubov vacuum also cannot capture all strong non-Gaussian correlations that may become important at very large hopping or close to the stability boundary. The widening lower–upper bracket in these regions is therefore a useful diagnostic that more elaborate correlated states or methods such as DMRG may be needed[30, 33]. It is not a failure of rigor: the Lieb–Yamazaki lower bound remains a strict lower estimate throughout the stable region, while the independently minimized Gaussian expectation is a variational upper estimate. Figure 1–Figure 3 likewise place the ED data inside the reported brackets. For example, evaluating the upper energy density at the lower-bound optimum gives, at , , at . This same-pair value is regulator dependent near the stability boundary. Independent upper minimization used in the present map gives . The reduction measures how rigidly the Gaussian state tied to the lower-bound optimum brackets the energy versus what the same Gaussian family can achieve under a separate variational minimization; the residual width diagnoses limitations of that ansatz rather than a numerical or mathematical inconsistency.
Convergence with the Local Occupation Cutoff
We finally tested the local Fock-space truncation used in the finite-chain ED data. Figure 8 follows the ground-state energy density through in four representative regimes. Using as the reference, the absolute error remains below at all subsequent sampled cutoffs from , 10, 10, and 12 in cases A–D, respectively. The corresponding cutoffs for the stricter test are 6, 10, 12, and 14. Thus the default is amply converged for the weak, strong-drive, and strong-hopping examples. It is not uniformly adequate in the mixed strong regime: for case D its error at is , and (or 14 for accuracy) should be used there. The criterion used throughout is therefore an absolute energy-density change below between consecutive adaptive cutoffs, supplemented by the explicit convergence study against ; no unverified relative criterion is assumed.
Figure 8. Convergence of the ED ground-state energy density with the local occupation cutoff. The panels show (a) the weak case , (b) strong drive , (c) strong hopping , and (d) the mixed strong case . Insets display on a logarithmic scale; dashed and dotted horizontal lines mark and , respectively.
Conclusions
A family of rigorous lower bounds has been derived for the ground-state energy of the coherently driven Bose–Hubbard model on periodic hypercubic lattices. By the positive-operator decomposition, followed by displacement and Bogoliubov diagonalization, the lower-bound energy density is obtained and its thermodynamic limit is established for every . By the same construction there is produced the continuous band of Proposition III.2, including its gap, effective mass, and boundary slope (an effective sound velocity of the excitations of the bounding Hamiltonian , not of the full driven model). Within the ground state of the bounding Hamiltonian, the order parameter, filling, energy densities, number variance, , and anomalous correlators are all explicit; in particular, the formulas identify the regime of drive-induced antibunching and sub-Poissonian fluctuations.
Finite-volume optimization for the ring changes either optimized bound from its thermodynamic-limit counterpart by at most approximately over the reported scan. The dominant width of the energy bracket therefore reflects the limitations of the Gaussian ansatz rather than finite-size effects. The independently optimized upper bound and the regime maps further locate where this bracket widens and trace the crossover curves separating antibunched and bunched correlations. Convergence of the ED markers is parameter dependent: meets the criterion in the weak and axis-strong regimes, whereas mixed strong regimes require adaptive local cutoffs starting at and extending through when necessary.
All reported observables, including , the excitation spectrum , and the boundary slope , are rigorous properties of the ground state of the auxiliary bounding Hamiltonian . They are not established observables or ground-state properties of the full driven Hamiltonian . Establishing whether the auxiliary gapless boundary has a counterpart in the full model requires independent finite-size scaling or tensor-network calculations such as DMRG or iTEBD.
The accuracy attainable with this quadratic reference state is also delimited by the numerical results. For the representative parameters and , the same-pair separation is about 31.2% of the lower-bound magnitude at , although this same-pair number is regulator dependent near a gapless stability boundary; the independently optimized upper bound is the more robust numerical comparator. Lemma V.1 makes the Temple test rigorous in finite volume, but its applicability condition reads . Thus an extensive trial-state energy excess eventually overwhelms the intensive certified gap. Temple’s inequality remains a conditional finite-size refinement, not a general thermodynamic correction or a test of whether the ansatz controls the full spectrum.
The elementary site and Anderson bond bounds added in § 7 A are sharper at several representative points, which sets a clear benchmark for the present analytic construction. Natural extensions include finite-temperature bounds, non-hypercubic or inhomogeneous geometries, and bounding states that incorporate non-Gaussian or higher-order correlations. The exactly soluble is thus retained as a Gaussian reference par excellence, rather than as a surrogate for the full interacting theory. Enlarging the positive-operator family or combining it with systematically improved variational states may increase the certified excitation threshold and reduce the energy variance, the two quantities that govern whether the finite-volume Temple condition can hold.
Acknowledgments
OpenAI Codex CLI and Qwen were used as language and coding assistance tools during the preparation and revision of this manuscript. All suggestions and generated material were critically reviewed by the authors, who bear full responsibility for the content of the work.
M.J. would like to dedicate this work to the memory of his Mother.
References
- [1]M. P. A. Fisher, P. B. Weichman, G. Grinstein, and D. S. Fisher, Physical Review B 40, 546 (1989).
- [2]D. Jaksch, C. Bruder, J. I. Cirac, C. W. Gardiner, and P. Zoller, Physical Review Letters 81, 3108 (1998).
- [3]M. Greiner, O. Mandel, T. Esslinger, T. W. Hänsch, and I. Bloch, Nature 415, 39 (2002).
- [4]I. Carusotto and C. Ciuti, Reviews of Modern Physics 85, 299 (2013).
- [5]M. J. Hartmann, Journal of Optics 18, 104005 (2016).
- [6]C. Noh and D. G. Angelakis, Reports on Progress in Physics 80, 016401 (2017).
- [7]J. Bloch, I. Carusotto, and M. Wouters, Nature Reviews Physics 4, 470 (2022).
- [8]T. Fink, A. Schade, S. Höfling, C. Schneider, and A. Imamoglu, Nature Physics 14, 365 (2018).
- [9]R. Ma, B. Saxberg, C. Owens, N. Leung, Y. Lu, J. Simon, and D. I. Schuster, Nature 566, 51 (2019).
- [10]F. Vicentini, F. Minganti, R. Rota, G. Orso, and C. Ciuti, Physical Review A 97, 013853 (2018).
- [11]T. Grab, Physical Review A 99, 043607 (2019).
- [12]D. Dahan, G. Arwas, and E. Grosfeld, npj Quantum Information 8, 14 (2022).
- [13]L. Rassaert, T. Ramos, T. Roscilde, and D. Porras, Physical Review Letters 135, 203603 (2025).
- [14]M. Lewin, P. T. Nam, S. Serfaty, and J. P. Solovej, Communications on Pure and Applied Mathematics 68, 413 (2015).
- [15]M. Napiórkowski, R. Reuvers, and J. P. Solovej, Archive for Rational Mechanics and Analysis 229, 1037 (2018).
- [16]B. Brietzke, S. r. Fournais, and J. P. Solovej, Communications in Mathematical Physics 376, 323 (2020).
- [17]P. W. Anderson, Physical Review 83, 1260 (1951).
- [18]E. H. Lieb and K. Yamazaki, Physical Review 111, 728 (1958).
- [19]N. N. Bogoliubov, Journal of Physics (USSR) 11, 23 (1947).
- [20]J.-P. Blaizot and G. Ripka, Quantum Theory of Finite Systems (MIT Press, 1986).
- [22]M. Reed and B. Simon, Methods of Modern Mathematical Physics. IV: Analysis of Operators (Academic Press, 1978).
- [24]S. Ghosh and T. C. H. Liew, Physical Review Letters 123, 013602 (2019).
- [25]R. Trivedi, M. Radulaski, K. A. Fischer, S. Fan, and J. Vučković, Physical Review Letters 122, 243602 (2019).
- [26]A. Kowalewska-Kudłaszyk, S. I. Abo, G. Chimczak, J. Perina, Jan, F. Nori, and A. Miranowicz, Physical Review A 100, 053857 (2019).
- [27]C. Menotti and N. Trivedi, Physical Review B 77, 235120 (2008).
- [28]M. Endres, M. Cheneau, T. Fukuhara, C. Weitenberg, P. Schauß, C. Gross, L. Mazza, M.-C. Baňuls, L. Pollet, I. Bloch, and S. Kuhr, Nature 487, 454 (2012).
- [29]S. R. White, Physical Review Letters 69, 2863 (1992).
- [30]U. Schollwöck, Annals of Physics 326, 96 (2011).
- [31]R. Roth, Annals of Physics 349, 117 (2014).
- [32]J. Hauschild and F. Pollmann, SciPost Physics Lecture Notes , 5 (2018).
- [33]S. Paeckel, T. Köhler, A. Swoboda, S. R. Mamnana, U. Schollwöck, and C. Hubig, Annals of Physics 411, 167998 (2019).







