PHYSICAL REVIEW B 83, 054508 (2011)

Enhanced pairing in the checkerboard Hubbard ladder

George Karakonstantakis,1 Erez Berg,2 Steven R. White,3 and Steven A. Kivelson3

1
Department of Physics, Stanford University, Stanford, California 94305, USA
2
Department of Physics, Harvard University, Cambridge, Massachusetts 02138, USA
3
Department of Physics and Astronomy, University of California, Irvine, California 92717, USA
(Received 23 August 2010; published 22 February 2011)
We study signatures of superconductivity in a two-leg “checkerboard” Hubbard ladder model, defined as a
one-dimensional (period 2) array of square plaquettes with an intraplaquette hopping t and interplaquette hopping
t  , using the density-matrix renormalization-group method. The highest pairing scale (characterized by the spin
gap or the pair binding energy, extrapolated to the thermodynamic limit) is found for doping levels close to
half-filling, U ≈ 6t and t  /t ≈ 0.6. Other forms of modulated hopping parameters, with periods of either one
or three lattice constants, are also found to enhance pairing relative to the uniform two-leg ladder, although to
a lesser degree. A calculation of the phase stiffness of the ladder reveals that in the regime with the strongest
pairing, the energy scale associated with phase ordering is comparable to the pairing scale.

DOI: 10.1103/PhysRevB.83.054508 PACS number(s): 74.20.−z

I. INTRODUCTION unit cell shown in Fig. 1(b), the optimal conditions occur for
an intermediate degree of inhomogeneity, t  /t ∼ 0.6 − 0.7,
The much debated theoretical issues related to the “mecha-
where t is the hopping matrix within a square and t  is the
nism” (i.e., microscopic origin) of high-temperature supercon-
hopping matrix between squares. This tends to corroborate
ductivity are often ill-defined. One related question to which
inferences made previously on the basis of exact diagonaliza-
unambiguous answers are possible is as follows: For a given
tion studies5 of the 2D “checkerboard-Hubbard model.” (iii) A
class of models, what values of the parameters are optimal for
qualitatively similar enhancement of superconductivity is ob-
superconductivity? Of course, if one can make predictions
served for the other periodic versions of the model with two or
about models, the same insights might provide guidance
six sites per unit cell, shown in Figs. 1(a) and 1(c), respectively,
in the search for materials with improved superconducting
although in these cases the magnitude of the effect is smaller
properties. Two specific questions we would like to address are
and the optimal condition occurs with values of t  /t closer to 1.
the following: (i) In the case in which superconductivity arises
The observation that certain patterns of spatial symmetry
directly from the repulsive interactions between electrons,
breaking can coexist with superconductivity (or even strongly
how strong (in units of the bandwidth) are the optimal
enhance it), while others do not, is also reminiscent of
interactions for superconductivity? (ii) Is there an “optimal
recent results obtained using DMRG7 and the dynamic
inhomogeneity” for superconductivity,1 in the sense of a
cluster approximation.8 In the first of these calculations, the
complex (but still periodic) electronic structure with multiple
inhomogeneity (in the form of stripes) occurs spontaneously,
orbitals per unit cell? An obvious difficulty with this program
while in the second it is imposed externally. As we were
is that, in most cases, we do not know how to compute the
completing this work, a contractor renormalization (CORE)
transition temperature of the relevant models in a controlled
study of the checkerboard Hubbard model in a 2D geometry
manner, so as to test the predictions of theory.
was presented in Ref. 9, extending earlier CORE results for
In this context, we use the density-matrix renormalization
the uniform 2D Hubbard model.10 Finite-size effects were
group (DMRG)2 to numerically compute the superconducting
found to be large for t   0.8t, but in the smaller t  regime,
correlations of the two-leg Hubbard ladder (extrapolated to
where these effects are relatively small, the results of this new
infinite length) as a function of the strength of the Hubbard
study are completely consistent with those of the earlier exact
interaction, U , and for various periodic patterns of the hopping
diagonalization studies,5 and lead to conclusions concerning
matrix elements. The one-dimensional (1D) character of the
the optimal conditions for superconductivity that are similar to
system studied is what permits us to obtain an accurate solution
those obtained in the present ladder study. The CORE method
of this problem. However, the same 1D character implies that
was also used to study ladders, albeit considerably shorter
no nonzero critical temperature is possible, so in assessing
than those studied in the present paper, and again the results
the optimal conditions for superconductivity, we are forced to
obtained are fully consistent with the present results.
use other energy scales in the problem, especially the spin gap,
Es , the pair-binding energy, Ep , and the superfluid helicity II. THE MODEL
modulus, ρc . We find the following: (i) The optimal value of
U is generally U ≈ 6t, where 6t is the total bandwidth of the We consider the repulsive U Hubbard model defined on a
uniform ladder. This result agrees with previous studies3,4 of (spatially modulated) two-leg ladder,
various ladder systems. It is also consistent with inferences  †
H=− (tj,j +1 cj,λ,σ cj +1,λ,σ + H.c.)
made on the basis of exact diagonalization5 and dynamical
j,λ,σ
cluster quantum Monte-Carlo6 studies of the 2D Hubbard  † 
model, where U ≈ 8t (i.e., the 2D bandwidth) was found to be −t (cj,1,σ cj,2,σ + H.c.) + U nj,λ,↑ nj,λ,↓. (1)
optimal. (ii) For the checkerboard pattern with four sites per j,σ j,λ

1098-0121/2011/83(5)/054508(8) 054508-1 © 2011 American Physical Society

KARAKONSTANTAKIS, BERG, WHITE, AND KIVELSON PHYSICAL REVIEW B 83, 054508 (2011)

may be multiple scales (e.g., multiple spin gaps) associated

with the gapped modes. However, χ0 remains a monotonic,
approximately linearly increasing function of Es .
Perhaps not surprisingly, we will see that the inhomoge-
neous Hubbard ladders we have studied are also Luther-Emery
liquids with Kc > 1/2. Thus, in addressing the “mechanism
of superconductivity,” the primary purpose of our DMRG
calculations is to determine the dependence of vc , Kc , Es ,
and ξs on microscopic parameters.
The pair binding energy Ep corresponds to creating
two spatially separated spin-1/2 quasiparticles. Since the
spins for these quasiparticles can either add to S = 0 or
1, we must have Es  Ep . If the residual interactions
FIG. 1. (Color online) Schematic representation of the “inhomo- between quasiparticles are repulsive, we expect Es = Ep .
geneous” Hubbard ladders considered in the present paper: (a) The Conversely, if the interactions between quasiparticles are
period-1 “dimer” ladder; (b) the period-2 “checkerboard” ladder; (c) attractive, a neutral spin-1 “exciton” is formed, which has
the period-3 ladder. As discussed below Eq. (1), the solid and dashed
lower energy than two far-separated quasiparticles, and hence
lines represent, respectively, hopping matrix elements t and t  .
Es < Ep . The latter behavior has been found previously
† in DMRG calculations on the uniform t − J ladder.13
Here cj,λ,σ creates an electron on rung j = 1, . . . ,L − 1 of
chain λ = 1,2 with spin polarization σ = ±, L is the length of
the ladder, U > 0 is the repulsion between two electrons on the IV. Energy scales
†
same site, the density operator is nj,λ,σ = cj,λ,σ cj,λ,σ , and n =
 The spin gap, Es , is the difference between the ground-
(2L)−1 j,λ,σ nj,λ,σ  is the mean number of electrons per site. state energies of the system with spin S = 1 and 0:
The much studied homogeneous Hubbard ladder corresponds
to the case tj,j +1 = t  for all j , although it is worth noting Es ≡ E0 (S = 1,2N ) − E0 (S = 0,2N ), (4)
that for t  t, this model can also be viewed as a coupled
array of Hubbard dimers. The “dimer ladder” is shown in where E0 (S,N ) is the spin S ground-state energy of the N-
Fig. 1(a). The “checkerboard ladder” in Fig. 1(b) has t2j,2j +1 = electron system.
t and t2j +1,2j +2 = t  < t. The “period-3” ladder in Fig. 1(c) Similarly, the pair-binding energy, Ep , is defined as
has t3j,3j +1 = t3j +1,3j +2 = t and t3j +2,3j +3 = t  < t.
The thermodynamic limit is accessed by computing quan- 1
Ep = 2E0 ,2N + 1 − E0 (0,2N ) − E0 (0,2N + 2). (5)
tities for various lengths, and then using finite-size scaling 2
analysis to extrapolate to 1/L → 0.
Were we computing these quantities in a BCS superconductor,
then in the thermodynamic limit, both these energies would be
III. EFFECTIVE FIELD THEORY equal to twice the minimum gap min ,
The uniform two-leg Hubbard ladder with n = 1 but still
not too far from n = 1 is well known on the basis of weak- lim Es = lim Ep = 2min .
L→∞ L→∞
coupling RG,11 bosonzation,12 and DMRG3 approaches to be
in a Luther-Emery phase characterized at low energies by a What we have in mind here is a system with a strongly k-
spin gap, Es [defined later in Eq. (4)] and a single, gapless dependent superconducting gap. In 2D, then, the value of the
acoustic “charge” mode that propagates with speed vc , and gap would depend on the position on the Fermi surface. For
whose long-range (power-law) correlations are determined by a ladder with a finite number of legs, there is a discrete set
a single Luttinger parameter, Kc . The Luther-Emery liquid can of transverse values of k, so even in the thermodynamic limit,
be thought of as a 1D version of a superconducting state in only certain discrete points on what would, in 2D, be a full
the sense that it has a nonvanishing superfluid stiffness [see Fermi surface are crossed. In this case, in the thermodynamic
Eq. (6)], and, for Kc > 1/2 and T Es , it has a divergent limit, the gap we would obtain would be the gap that occurs
superconducting susceptibility, at the point on the 2D Fermi surface where the gap happens
 v (2−1/Kc ) to be smallest. For an s-wave superconductor (SC), this is
c
χ ∼ χ0 , (2) a reasonable measure of the gap in 2D. For a d-wave SC,
aT
the precise value depends on how close the closest Fermi
where vc is the charge velocity and a is a lattice constant. In surface crossing is to the nodal point. Thus, it is intuitively
the single chain realization of a Luther-Emery liquid, reasonable to associate these energy scales with a mean-field
      estimate of the superconducting critical temperature, TcMF ∼
a aEs a a
χ0 = = , (3) Es /4. Of course, since the ladder is a 1D system, the actual
vc vs vc ξs
Tc = 0.
where ξs = vs /Es is the spin-correlation length and vs is the While it may be reasonable to interpret Es and/or Ep
spin velocity. For a multicomponent system, the corresponding as measures of a pairing scale in the problem, to address
expression for χ0 is somewhat more complicated, as there the growth of superconducting correlations it is ultimately

054508-2
ENHANCED PAIRING IN THE CHECKERBOARD HUBBARD . . . PHYSICAL REVIEW B 83, 054508 (2011)

necessary to consider the helicity modulus, which governs the as a function of distance down the ladder. In a spin-gapped
energetics of superconducting phase fluctuations: phase, we expect
 † †
M(j ) = σ [cj,1,σ cj,1,σ − cj,2,σ cj,2,σ ]
vc Kc ∂ 2 E0
ρc = ≡ lim L 2 , (6) σ
2π L→∞ ∂φ φ=0 ∼ cos[Qj + φ0 ] exp[−|j |a/ξs ]. (8)
In the limit of an asymptotically small spin gap, Q = 2kF , but
where, in this case, the ground-state energy is computed in the for larger gaps it may depend not only on n but on U/t as well.
presence of pair fields applied to the two ends of the system To be explicit, we therefore define the spin-correlation length
with a relative phase twist φ. as
In 2D, the relative importance of phase and pair-breaking 
fluctuations can be assessed14 by considering the ratio of the j |j M(j )|
ξs =  . (9)
phase stiffness (which has units of energy) to the pairing gap. j |M(j )|
However, in 1D, ρc has units of a velocity, so defining an energy
It turns out that Eq. (6) is relatively difficult to implement to
scale, Eθ , characteristic of the phase fluctuations requires
obtain quantitatively reliable results for ρc using DMRG. How-
introducing a length scale in the problem. The important
ever, it is possible15 to compute ρc by separately calculating
(longest) emergent length scale is ξs , in terms of which we
vc and vc /Kc from quantities that are more straightforwardly
define
computed using DMRG. From the bosonized field theory, we
can identify the inverse compressibility of the ladder with the
Eθ ≡ πρc /ξs ≡ R Es . (7) ratio 2Kπvc
. In turn, in all circumstances relevant to the present
c
calculation,16 the compressibility is related to the energy to
Here R ≡ Eθ /Es is the dimensionless ratio of the phase add or remove a singlet pair of electrons from the ladder:
ordering and pairing scales.
1 E0 (0,2N + 2) + E0 (0,2N − 2) − 2E0 (0,2N )
To appreciate the significance of this ratio, consider its = lim L .
value for the attractive Hubbard chain in various limits. The κ L→∞ 4
1D version of a BCS limit, in which there is a single character- (10)
istic energy and temperature scale, s ∼ exp[−π vs /a|U |], An independent measurement of vc can be obtained by
is realized in the limit |U | 1 where, up to corrections calculating also the energy of the first excited state,
of order U/t, vs = vc and Kc = 1, so R = vc Kc /2vs = E1 (S,N ),according to
1/2 + O(U/t), that is, both mesoscale phase coherence and
pairing correlations onset at a temperature of the order of L
vc = lim [E1 (0,2N ) − E0 (0,2N )]. (11)
Tpair ∼ Es /4. Conversely, R → 0 as |U |/t → ∞; for large L→∞ π
U , a spin pseudogap opens when T ∼ Tpair = |U |/2, with a We then compute the helicity modulus as
second crossover from a largely incoherent paired state to
a coherent Luther-Emery liquid occurring at a temperature vc2 κ
Tθ ∼ Eθ ∝ t 2 /|U |, well below Tpair . A similar dichotomy ρc = . (12)
4
exists in the two-leg repulsive U Hubbard ladder, where
Note that this procedure also gives us
R → 0 as the doped hole concentration, x → 0, while R ∼ 1
at larger values of x where the spin gap is significantly π
Kc = κvc . (13)
suppressed relative to its value at x = 0. In the small-x case, 2
the doped holes can be thought of as a dilute gas of charge An alternative way to obtain Kc is by measuring the amplitude
2e bosons at temperatures small compared to Tpair , but the of the “Friedel-like” density oscillations, which exhibit a
phase-coherence scale is much smaller and vanishes as x → 0. power-law decay as a function of distance from the edge of
With these examples in mind, we identify the case R ∼ 1 the system. For long systems, the density near the center of a
with the 1D version of the “BCS-like limit” in which there length L ladder takes the form15
is a single crossover temperature Tpair that separates the
“normal” (multicomponent Luttinger liquid) high-temperature cos[2π n(j − L/2)]
nj  ∼ . (14)
regime from the low-temperature regime in which substantial LKc /2
mesoscale superconducting order has developed. Conversely, Therefore, by measuring the amplitude of the density oscilla-
if R 1, two distinct crossover scales characterize the evolu- tions ACDW versus L and plotting ln(ACDW ) versus ln(L), Kc
tion from the normal state: a first, high-temperature crossover, can be obtained. Whenever possible, we have calculated Kc
Tpair , characterized by the opening of a spin pseudogap, and using both Eqs. (13) and (14), and we found that the two values
a lower crossover temperature, Tθ ∼ Eθ /4, which can be agree with each other to within about 10%.
viewed as the scale at which the liquid of bosonic pairs begins
to exhibit substantial local phase coherence.
V. DMRG RESULTS
The most direct and efficient way to compute ξs from
DMRG is to apply a staggered Zeeman field to one end We have computed ground-state properties for ladder
of the ladder, j = 0 (thus locally breaking spin-rotational systems for various values of n, t  /t, and U/t using DMRG.
symmetry), and then measure the decay of the magnetization We have kept up to m = 2400 states and extrapolated our

054508-3
KARAKONSTANTAKIS, BERG, WHITE, AND KIVELSON PHYSICAL REVIEW B 83, 054508 (2011)

<n>=0.875
<n>=0.75
<n>=0.9375

0.1
ΔΕs(units of t)

0.05

0
0 0.2 0.4 0.6 0.8 1
t′ (in units of t)

FIG. 2. (Color online) The spin gap, Es , of the checkerboard

Hubbard ladder as a function of t  /t for n = 0.9375, 0.875, and 0.75
at fixed U = 8t, extrapolated to the thermodynamic limit (L → ∞). FIG. 4. (Color online) The pair-binding energy, Ep , of the
checkerboard Hubbard ladder for n = 0.875 (in the thermodynamic
limit) as a function of U and t  , fixing t = 1.
results to zero truncation errors. As is well known,17 ground-
state energies (as well as one-point correlation functions18 )
can be extracted with great accuracy in this way. Results have 4 − 8t) and intermediate inhomogeneity, t  /t ∼ 0.5. This
been obtained for system sizes from 2 × 16 up to 2 × 64, figure looks qualitatively similar to the analogous result for
and then extrapolated to the thermodynamic limit (1/L → 0) the two-dimensional checkerboard Hubbard model obtained
using a finite-size scaling analysis. For an example of this previously by exact diagonalization of a 16-site system in
procedure, see Appendix A. Since DMRG converges better Ref. 5; however, in contrast to that study, the present results
for open boundary conditions, all the calculations were done are obtained in the thermodynamic limit.
using open boundary conditions in the long direction. From The dependence of Ep on U/t and t  /t is generally
the extrapolated values, we have extracted Es , ξs , Ep , ρc , similar to that of Es , as can be seen by comparing the
and Kc , as described earlier. contour plots of these two quantities for n = 0.875, which
In Fig. 2, we show Es for fixed U/t = 8 as a function are shown in Figs. 4 and 3, respectively. However, there are
of t  /t for n = 0.9375, 0.875, and 0.75. Note that the value interesting differences, as can be seen in Fig. 5, where the two
of Es rises from its value for the uniform ladder as t  /t quantities are plotted as a function of t  /t for fixed U/t = 8
is reduced below t  /t = 1, reaches a maximum value at an and n = 0.875. Note that for t  /t > 0.6, Ep > Es . This is,
intermediate value of t  /t, and then drops to zero as t  /t → 0. presumably, indicative of the existence of a spin-1 excitonic
For instance, for n = 0.875, the maximum value Es ≈ 0.12t, bound state for t  /t > 0.6. A similar result was found in the
which occurs for t  /t = 0.6, is approximately four times larger uniform two-leg t-J model ladders.13,19
than its value in the uniform ladder. More broadly, we have To calculate R = Eθ /Es , we must compute ρc and ξs .
studied the spin gap as a function of both U/t and t  /t; the To obtain ξs , we apply a relatively strong staggered Zeeman
results for n = 0.875 are shown in Fig. 3. One can see that Es field of magnitude t to the end sites of the ladder and measure
exhibits a broad maximum for U of order the bandwidth (U ∼ the decay of the staggered magnetization, M(j ), as in Eq. (8).
In all cases, we have found that M(j ) decays rapidly on scales
short compared to the length of our longest ladders, so ξs can
be extracted from the calculations accurately. Representative
results for M(j ) are shown in the inset of Fig. 6. ξs as a function

0.2
ΔE
s
0.15 ΔE
p

0.1

0.05

0
0.2 0.4 0.6 0.8 1
t′/t

FIG. 5. (Color online) Es (L → ∞) and Ep (L → ∞) for the

FIG. 3. (Color online) Es (L → ∞) of the checkerboard Hub- checkerboard ladder with fixed n = 0.875 and U/t = 8, as a function
bard ladder for n = 0.875 as a function of U and t  , fixing t = 1. t  /t.

054508-4
ENHANCED PAIRING IN THE CHECKERBOARD HUBBARD . . . PHYSICAL REVIEW B 83, 054508 (2011)

15 TABLE I. Values of the ratio R defined in Eq. (7) for n = 0.875

0.5
and U = 8t.

〈Sz (x)〉
10
0
t  = 0.2 0.4 0.6 0.7 0.8 1.0
R = 3.38 3.06 0.96 1.01 0.99 0.98
ξs /a

−0.5
0 20 40 60
x
5
Thus, at least crudely, this regime can be thought of as a
“BCS-like” regime, in which there is a single energy scale,
0
0 0.2 0.4 0.6 0.8 1 set by Es , that characterizes the growth of superconducting
t′/t correlations. Depending on precisely what criterion one
chooses to quantify the crossover scale, phase fluctuations
FIG. 6. (Color online) The spin correlation length, ξs , for the
checkerboard ladder with U = 8t and n = 0.875 as a function of
will produce a quantitative difference in the magnitude of the
t  /t, calculated from Eq. (9). The inset shows the expectation value specified scale, but not large qualitative effects. Therefore, it
of the spin S z  for U = 8t, n = 0.875, and t  /t = 1, as a function of is reasonable to assert that the values of the parameters that
position. A staggered Zeeman field of strength t has been applied to lead to the largest values of Es and/or Ep are the “optimal
the two sites at the left edge of the ladder. values for superconductivity.”
For t  < 0.5t, we obtain R ∼ 3, that is, Eθ > Es ,
of t  /t is shown in Fig. 6, for fixed n = 0.875 and U/t = 8. suggesting that this regime cannot be thought of in terms of
Note that for t  /t < 1/2, the spin-correlation length is roughly either a naive weak- or strong-coupling picture. Remarkably,
3a, which is of the order of one unit cell of the checkerboard the transition from R ∼ 1 to R > 3 occurs quite sharply around
ladder. t  = 0.5, close to the point where the spin gap is optimal.
Next, we calculate both ρc and Kc following the procedure It is interesting to note that for n = 0.75, t  /t = 0.4, we
described above [Eqs. (10)–(14)]. The value of Kc is shown find a sharp decrease of Kc and ρc . The value of Kc at this
in Fig. 7 for n = 0.75, 0.875, and 0.9375, fixing U/t = 8, point is smaller than the critical value of 1 at which a static
as a function of t  /t. In contrast to the results for Es charge-density wave should be stable,15 indicating that this
(and somewhat to our surprise), for n = 0.875, Kc is a behavior of Kc and ρc may be due to a charge-density-wave
weakly varying function of t  /t (and, as it turns out, U/t phase that exists for n = 0.75, t   0.4t.
is as well). To a good approximation, for a wide range We thus conclude that for the checkerboard Hubbard ladder,
of values, we can simply take Kc ≈ 1, independent of t  /t optimal superconductivity occurs for intermediate values of
and U/t. Note that this implies that the superconducting U/t ∼ 6, intermediate inhomogeneity, t  /t ∼ 0.6 − 0.7, and
susceptibility diverges as T → 0, so that it is reasonable to electron concentrations near (but not equal to) one electron per
think of the ladder as a fluctuating superconductor. (Of course, site. We can now ask if this result is special to the checkerboard
there is also a divergent charge-density-wave susceptibility, pattern, or if it applies more generally to the situation in which
χCDW ∼ T −(2−Kc ) , so there is some unavoidable ambiguity there are multiple sites per unit cell. We thus have repeated
with this simple intuitive picture.) As n is increased to 0.9375, (although not in as much detail) the same calculations for the
Kc increases, consistent with the expectation that Kc → 2 as dimer ladder (period 1) and the period-3 ladder. (See Fig. 1.)
n → 1.20 In Fig. 8, we exhibit the dependence of the spin gap of all
From the measured values of ξs , κ, and Kc , the energy three ladders for fixed U = 8t and n = 0.875 as a function of
scale characteristic of phase-ordering can be extracted. Table I t  /t. In all the cases, we see that there is an increase in the spin
shows the ratio R from Eq. (7). Note that for t  /t > 0.5, R ≈ 1. gap for some t  /t < 1.
The result was expected, qualitatively, in the dimer (period-
1) case from previous works,3,4,21 which found that the spin
1.6 0.25
gap (as well as pairing correlations) is enhanced upon making
n=0.9375 0.2
t  smaller than t in the dimer (period-2) ladder. In the case of
1.4 the period-3 ladders, there is a very weak increase of the spin
0.75 0.15 gap, which occurs at t  /t ∼ 0.9.
n=0.75
c

In Ref. 4, it was argued that the enhancement of super-

K

1.2
ρ

0.875 0.875
0.1 conducting correlations in the dimer ladder is due to the
0.9375
1 increase in the density of states close to the “Van Hove”
0.05
point, in which one of the two bands of the two-leg ladder
0.8 0 becomes unoccupied. Beyond this point, there is only a single
0.2 0.4 0.6 0.8 1 0 0.5 1
band crossing the Fermi level, and the system is likely to
t′ t′
behave as a single-component Luttinger liquid. Therefore, the
FIG. 7. (Color online) Left: the Luttinger parameter Kc as a superconducting signatures are rapidly suppressed. Consistent
function of t  /t for n = 0.75, 0.875, and 0.9375, and U = 8t. The with this picture, in the dimer ladder we find that the spin
error bars were estimated by comparing between the values of Kc gap collapses to zero below t  /t  0.6. In the period-2 (the
obtained from Eqs. (13) and (14). Note that according to our definition checkerboard ladder) and period-3 cases, however, no such
of Kc , the noninteracting value is Kc = 2. Right: The phase stiffness sudden suppression of the spin gap is observed as t  /t is
ρc [defined in Eq. (12)] as a function of t  /t. reduced below the optimal point. This leads us to believe that

054508-5
KARAKONSTANTAKIS, BERG, WHITE, AND KIVELSON PHYSICAL REVIEW B 83, 054508 (2011)

The period 3 ladder

The checkerboard ladder
The dimer ladder
ΔΕ s (in units of t)

0.1

0.05

FIG. 9. (Color online) A system of coupled checkerboard ladders,

connected by a single particle tunneling matrix element t  .

0
results for the checkerboard Hubbard
ladder with n = 0.875
0 0.2 0.4 0.6 0.8 1 and U = 8t, we recall that Kc ( tt ) ≈ 1 over the entire range
t′ (in units of t) 0 < t   1 (see Fig. 7). We therefore fix Kc = 1, independent
FIG. 8. (Color online) Es (L → ∞) for the three types of
of t  . The resulting expression for Tc is
√   2
inhomogeneous ladders in Fig. 1 is shown at fixed n = 0.875 and K( 1 − x 2 ) at
U/t = 8 as a function of t  /t. The inhomogeneity induced by Tc ∼ Es . (15)
breaking up the ladders to period-1, -2, and -3 clusters increases
x vc
the spin gap for t  /t < 1. The increase is most dramatic for the Here, x ≡ vs /vc , and K(x) is a complete elliptic integral of the
checkerboard ladder, in which the maximum spin gap is about four first kind. Note that Tc depends on t  /t through vs , vc , and Es .
times larger than the spin gap for the uniform (t  = 1) system. For the As t  decreases, both vc and vs decrease; their ratio, however, is
period-1 (dimer) ladder, the enhancement is by a factor of 2, while for found to be approximately constant as a function of t  /t down
the period-3 ladder, the spin gap is only slightly enhanced (by about to about t  /t = 0.5. (The value of vs is obtained by using
10%). the estimate Es ξs , where both Es and ξs are calculated
the mechanism of the enhancement of the spin gap for t  < t in from DMRG.) Es ( tt ), on the other hand, has a maximum for
t 
the checkerboard and period-3 ladders is unlikely to be related t
< 1. Therefore, as tt is reduced from 1, Tc (t  ) necessarily

to a proximity to a Van Hove point. increases, and reaches a maximum for some tmax < t.
Note also that for all the inhomogeneous patterns in Fig. 8,
the spin gap seems to reach zero at a critical tc > 0 (which is VII. DISCUSSION
different for each pattern). In particular, for the “checkerboard”
pattern, tc ∼ 0.05t. It is likely that for t  < tc , the Luther- The present study, along with a variety of other re-
Emery phase gives way to a Luttinger-liquid phase with one cent studies,5,8,9,22 provides strong support for a number of
gapless charge mode and one gapless spin mode (or more), intuitively appealing ideas concerning the physics of the
although more work is needed to establish that. superconducting Tc in unconventional superconductors in
Overall, among all the patterns we have reported, the which the pairing arises directly from the repulsive interactions
optimal ladder for superconductivity is a checkerboard ladder between electrons: (i) The highest superconducting transition
with U = 6t, t  /t = 0.6 − 0.8, and n = 0.875, for which temperatures occur at intermediate interaction strength, when
Es = 0.12t and Ep = 0.16t. U is comparable to the bandwidth. (A corollary of this is
that materials that are studied because of their high transition
temperatures are also likely to exhibit more general signatures
VI. EXTENSION TO QUASI-1D
of lying in an intermediate-coupling regime; here, theoretical
Earlier in the paper, we argued that the superconducting results from both weak- and strong-coupling approaches
tendency in the checkerboard Hubbard ladder is optimized for must be extrapolated, at best, to the limits of their regimes
an intermediate value of t  /t. However, since the supercon- of applicability.) (ii) Certain mesoscale structures (“optimal
ducting Tc of that system (as in any 1D system) is strictly inhomogeneity”1 ) can lead to enhanced superconducting
zero, one can worry that this statement may depend on how pairing, although clearly if the system is too strongly inhomo-
one chooses to measure the strength of the superconducting geneous, that always leads to a suppression of global phase
correlations. We will now consider a system composed of coherence. (iii) While short-range magnetic correlations,
an array of parallel checkerboard Hubbard ladders coupled possibly of the sort envisioned in the putative resonating
weakly in the direction transverse to the ladders, in which Tc valence bond state of a quantum antiferromagnet or in certain
can be estimated in a controlled way based on the solution of theories of a spin-fluctuation exchange mechanism, may well
the single-ladder problem. We will show that Tc is maximal for be important for pairing, longer-range magnetic correlations,
t
t
< 1. Thus, in this system, Tc is indeed optimized when the especially of the sort one would expect near a magnetic
electronic structure is nonuniform; that is, there is an “optimal quantum critical point, do not appear to be particularly
degree of inhomogeneity” for superconductivity. favorable for superconductivity. (This final conclusion follows
The quasi-1D system of coupled checkerboard Hubbard from a comparison of the t  /t dependence of the magnetic
ladders is depicted in Fig. 9. The ladders are coupled by a correlation length and the superconducting pairing in Figs. 6
single-particle tunneling matrix element t  . We fix the value and 4, respectively.)
of t  t,t  , and estimate Tc (t  /t) from an interchain mean- In addition, we found that the two-leg ladder at intermediate
field theory, described in Appendix B. From the numerical coupling (with U of the order of the bandwidth) and close to

054508-6
ENHANCED PAIRING IN THE CHECKERBOARD HUBBARD . . . PHYSICAL REVIEW B 83, 054508 (2011)

half-filling is, in many respects, surprisingly well described 0.14

t′=1
as a “BCS-like” superconductor, in which there is a single t′=0.8
crossover energy scale from the “normal” to the “supercon- 0.12
t′=0.6
ducting” state (rather than two separate scales, associated with 0.1 t′=0.4
pairing and phase coherence). This is based on the fact that t′=0.2
the ratio of the pairing and phase-coherence scales [defined in 0.08

s
ΔE
Eq. (7)] is close to its weak-coupling value, which justifies the
identification of the spin gap Es (or the pair-binding energy 0.06
Ep ) as the relevant energy scale for superconductivity. 0.04
Finally, there are a few unresolved issues and further
directions we would like to highlight: (i) The extrapolation 0.02
of the present results to higher dimensions is, of course, the
most important open issue. The strong qualitative similarity 0
0 0.02 0.04 0.06
between the present results and those obtained by exact 1/ L
diagonalization and CORE calculations on relatively small
2D clusters certainly encourages us to believe that the results FIG. 10. (Color online) Spin gap Es as a function of 1/L for
systems of sizes 2 × L, where L = 16,32,64, with n = 0.875, U/t =
obtained here give insight into the behavior of the higher-
8, and various values of t  /t. Symbols: DMRG results; solid lines:
dimensional problem. In this context, it might be useful to
second-order polynomial fits.
carry out similar calculations on four-leg and possibly even
six-leg ladders and cylinders, although it is considerably more smallest at values of t  /t that correspond to the maximum spin
difficult to extend these results to such long systems as are gap. We found that this behavior repeats itself for other values
accessible for the two-leg ladder. (ii) It is not clear exactly of n and U/t.
what aspects of the local electronic structure are essential
features of an optimal inhomogeneity for superconductivity. APPENDIX B: INTERCHAIN MEAN-FIELD THEORY
In the present case, it is notable that pair binding does not
occur on an isolated dimer or six-site rectangle for any value In this appendix, we describe the interchain mean-field
of U/t, while there is pair binding on an isolated square for treatment of the quasi-one-dimensional system described in
U/t < 4.6. However, this observation does not provide an Sec. VI. This procedure is quite standard.23–25 We consider
entirely satisfactory account of our findings, since the optimal an array of plaquette ladders, modeled by Luther-Emery
pairing in the checkerboard ladder occurs for U/t = 4 − 8t, liquids. For simplicity, we will assume that each ladder is a
where the pair-binding energy of an isolated square is either single-component system with a spin gap Es . (The extension
small or negative. to the case of a two-component system is straightforward, and
the result is qualitatively the same.) The ladders are coupled
ACKNOWLEDGMENTS by an interchain hopping term of the form
  †
We thank Ehud Altman, Assa Auerbach, Malcolm Beasley, H⊥ = −t  dxψP σ (x,n)ψP σ (x,n + 1), (B1)
Thierry Giamarchi, Lilach Goren, Dror Orgad, Didier Poil- σ,P =± n
blanc, Doug J. Scalapino, Alexei Tsvelik, and Wei-Feng Tsai
†
for useful discussions. This research was supported by the where ψP σ (x,n) (P = ±) creates a right- or left-moving
NSF under Grant No. DMR-0531196 (S.A.K.), and SIMES electron with spin σ =↑ , ↓ at position x in chain n. Next,
under Grants No. DMR-0197960 (G.K.), No. DMR-0907500 we integrate out degrees of freedom of length scales smaller
(S.R.W.), No. DMR-0705472, and No. DMR-0757145 (E.B.). than the spin correlation length ξs ∼ E vs
s
. Over such length
scales, the system is essentially gapless and can be treated as a
APPENDIX A: EXTRAPOLATION TO THE Luttinger liquid. To second order in t  , the following effective
THERMODYNAMIC LIMIT interchain action is generated:
The physical quantities reported in this paper are mostly  †
S⊥eff = (t  )2 dx dτ dx  dτ  T ψ+,σ (x,τ,n)
extrapolated to the thermodynamic (L → ∞) limit. This is
σ σ  ,n
done by calculating the corresponding quantity for various
†
system sizes (typically we have used L = 16,32, and 64 rungs) × ψ+,σ (x,τ,n + 1) ψ−,σ  (x  ,τ  ,n)
and then extrapolating to the limit 1/L → 0. As an example
× ψ−,σ  (x  ,τ  ,n + 1)0,> , (B2)
of this procedure, we present the spin gap Es for n = 0.875,
U = 8t, and various values of t  /t, as a function of 1/L, in where · · ·0,> denotes averaging over the “fast” (short-
Fig. 10. We use a second-order polynomial in 1/L to fit the wavelength) degrees of freedom [using the decoupled (t  = 0)
data and extrapolate, which in most cases fits the finite-size action], and T denotes time ordering. Since we are essentially
data well. performing a “coarse-graining” step, increasing the cutoff
Overall, the extrapolation to the thermodynamic limit gives of the theory from the lattice constant a to ξs , the region
a correction of up to about 30% to the measured values, of integration in Eq. (B2) is (x − x  )2 + vs2 (τ − τ  )2 < ξs .
making it the largest source of error in our calculations (DMRG To evaluate the integrand,√ we write the fermionic fields in
truncation errors are typically smaller than the symbol sizes in bosonized form: ψP σ ∼ ei π(θσ +P ϕσ ) , where ϕσ and θσ are dual
Fig. 10). Interestingly, the amount of extrapolation needed is bosonic fields that satisfy [ϕσ (x),∂x θσ  (x  )] = iδσ σ  δ(x − x  ).

054508-7
KARAKONSTANTAKIS, BERG, WHITE, AND KIVELSON PHYSICAL REVIEW B 83, 054508 (2011)

√ √
As usual, we introduce also √ charge and spin fields de- where (x,n) = ψR↑ ψL↓ − ψR↓ ψL↑ ∼ ei 2π θc
cos 2π ϕs
fined√as ϕc,s = (ϕ↑ ± ϕ↓ )/ 2, and similarly θc,s = (θ↑ ± and
θ↓ )/ 2. We define the fermionic Green’s function G(x,τ ) = ⎡ ⎤
 2   2
T ψ+,↑ (x,τ,n)ψ−,↓ (0,0,n)0,> . Expressing G(x,τ ) in terms of J⊥ ∼ K ⎣ 1 −
vs ⎦ at vc
. (B5)
the bosonic fields, vc vc a
 √ θc +θc θs −θs ϕc −ϕc ϕs +ϕs  √
G(x,τ ) ∼ ei 2π [ 2 + 2 + 2 + 2 ] 0,>  π/2
Here, K(α) = 0 dλ/ 1 − α 2 sin2 λ is a complete elliptic
1 Kc 1
a2 8Kc
a2 8
a2 4 integral of the first kind. Equation (B5) contains a Kc -
∼ 2 dependent prefactor, which we omit.
x + (vc τ )2 x 2 + (vc τ )2 x 2 + (vs τ )2 The mean-field equation for Tc is
√
2π[θ̄c +ϕ̄s ]
× ei , (B3)
zJ⊥ χ (Tc ) = 1, (B6)
where we have used the shorthand notation θc ≡ θc (x,τ ),
θc ≡ θc (x  ,τ  ), θ̄c ≡ θc (x/2,τ/2), and similarly for θs , ϕc , and where χ (T ) is the superconducting susceptibility of a single
ϕs . Plugging G(x,τ ) into Eq. (B2) and performing the integral, chain, and z is the number of nearest-neighbor chains (e.g.,
we get that the following interchain Josephson coupling for a two-dimensional array of checkerboard ladders, z = 2).
term: Inserting Eqs. (2) and (B5) in the mean-field equation (B6),
 and using the fact that for the checkerboard Hubbard ladder
H⊥eff = −J⊥ dx (x,n)† (x,n + 1) + H.c., (B4) Kc ≈ 1 over a wide range of parameters, we obtain Eq. (15)
n for Tc .

1 14
S. Kivelson and E. Fradkin, in Handbook of High-Temperature V. J. Emery and S. A. Kivelson, Nature (London) 374, 434
Superconductivity Theory and Experiment, edited by J. R. Schrieffer (1995).
15
and J. Brooks (Springer, New York, 2007), Chap. 15. S. R. White, I. Affleck, and D. J. Scalapino, Phys. Rev. B 65, 165122
2
S. R. White, Phys. Rev. Lett. 69, 2863 (1992). (2002).
3 16
R. Noack, S. White, and D. Scalapino, Physica C 270, 281 (1992). M.-S. Chang and I. Affleck, Phys. Rev. B 76, 054521 (2007).
4 17
R. M. Noack, N. Bulut, D. J. Scalapino, and M. G. Zacher, Phys. J. Bonca, J. E. Gubernatis, M. Guerrero, E. Jeckelmann, and S. R.
Rev. B 56, 7162 (1997). White, Phys. Rev. B 61, 3251 (2000).
5 18
W.-F. Tsai, H. Yao, A. Lauchli, and S. A. Kivelson, Phys. Rev. B S. R. White and A. L. Chernyshev, Phys. Rev. Lett. 99, 127004
77, 214502 (2008). (2007).
6 19
T. A. Maier, M. Jarrell, and D. J. Scalapino, Phys. Rev. B 74, 094513 D. Poilblanc, E. Orignac, S. White, and S. R. Capponi, Phys. Rev.
(2006). B 69, 220406 (2004).
7 20
S. R. White and D. J. Scalapino, Phys. Rev. B 79, 220504(R) (2009). H. J. Schulz, Phys. Rev. B 59, 2471(R) (1999).
8 21
T. A. Maier, G. Alvarez, M. Summers, and T. C. Schulthess, Phys. J. Riera, D. Poilblanc, and E. Dagotto, Eur. Phys. J. B 7, 53
Rev. Lett. 104, 247001 (2010). (1999).
9 22
S. Baruch and D. Orgad, Phys. Rev. B 82, 134537 (2010). H. Yao, W.-F. Tsai, and S. A. Kivelson, Phys. Rev. B 76, 161104(R)
10
E. Altman and A. Auerbach, Phys. Rev. B 65, 104508 (2002). (2007).
11 23
L. Balents and M. P. A. Fisher, Phys. Rev. B 53, 12133 (1996). E. Orignac and T. Giamarchi, Phys. Rev. B 56, 7167 (1997).
12 24
H. J. Schulz, Phys. Rev. B 53, 2959(R) (1996). T. Giamarchi and A. M. Tsvelik, Phys. Rev. B 59, 11398 (1999).
13 25
D. Poilblanc, O. Chiappa, J. Riera, S. R. White, and D. J. Scalapino, E. W. Carlson, D. Orgad, S. A. Kivelson, and V. J. Emery, Phys.
Phys. Rev. B 62, 14633(R) (2000). Rev. B 62, 3422 (2000).

054508-8