On the Dimension of Unimodular Discrete SpacesPart II: Relations with Growth Rate

# On the Dimension of Unimodular Discrete Spaces Part II: Relations with Growth Rate

François Baccelli111The University of Texas at Austin, baccelli@math.utexas.edu, Mir-Omid Haji-Mirsadeghi222Sharif University of Technology, mirsadeghi@sharif.ir, and Ali Khezeli333Tarbiat Modares University, khezeli@modares.ac.ir
July 15, 2019
###### Abstract

The notions of unimodular Minkowski and Hausdorff dimensions are defined in [5] for unimodular random discrete metric spaces. The present paper is focused on the connections between these notions and the polynomial growth rate of the underlying space. It is shown that bounding the dimension is closely related to finding suitable equivariant weight functions (i.e., measures) on the underlying discrete space. The main results are unimodular versions of the mass distribution principle and Billingsley’s lemma, which allow one to derive upper bounds on the unimodular Hausdorff dimension from the growth rate of suitable equivariant weight functions. Also, a unimodular version of Frostman’s lemma is provided, which shows that the upper bound given by the unimodular Billingsley lemma is sharp. These results allow one to compute or bound both types of unimodular dimensions in a large set of examples in the theory of point processes, unimodular random graphs, and self-similarity. Further results of independent interest are also presented, like a version of the max-flow min-cut theorem for unimodular one-ended trees.

## 1 Introduction

This paper is the second in a series of three ([5], the current paper, and [6]), which are referred to as Part I, II and III below. The present paper uses the definitions and symbols of Part I. A list of notation is provided in Table 1 at the end of the paper to ease the reading. For cross referencing the definitions and results of Part I, the prefix ‘I.’ is used. For example, Definition I.2.1 refers to Definition 2.1 in Part I.

Part I introduced the notion of unimodular random discrete metric space and two notions of dimension for such spaces, namely the unimodular Minkowski dimension (Section I.3.1 in Part I) and the unimodular Hausdorff dimension (Section I.3.3).

The present paper is centered on the connections between these dimensions and the growth rate of the space, which is the polynomial growth rate of , where represents the closed ball of radius centered at the origin and is the number of points in this ball.

Section 2 is focused on the basic properties of these connections. It is first shown that the upper and lower polynomial growth rates of (i.e., limsup and liminf of as ) provide upper and lower bound for the unimodular Hausdorff dimension, respectively. This is a discrete analogue of Billingsley’s lemma (see e.g., [9]). A discrete analogue of the mass distribution principle is then provided, which is useful to derive upper bounds on the unimodular Hausdorff dimension. In the particular case of a point-stationary point process equipped with the Euclidean metric, it is also shown that the unimodular Minkowski dimension is bounded from above by the polynomial decay rate of . Weighted versions of these inequalities, where a weight is assigned to each point, are also presented.

Section 3 is devoted to examples. It continues the main example section of Part II and introduces further examples. It is shown there that the bounds established in Section 2 are very useful for calculating the unimodular dimensions in the main instances of unimodular discrete spaces discussed in Part I, namely point processes, random graphs and self similar random discrete sets.

Section 4 gives a unimodular analogue of Frostman’s lemma. Roughly speaking, this lemma states that there is a weight function such that the upper bound in Billingsley’s lemma is sharp. This lemma provides a powerful tool to study the unimodular Hausdorff dimension, in particular, to assess the dimension of subspaces and product spaces. In the Euclidean case, another proof of the unimodular Frostman lemma is provided using a unimodular version of the max-flow min-cut theorem, which is of independent interest.

## 2 Connections to Growth Rate

Let be a discrete space and . The upper and lower (polynomial) growth rates of are

 ¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯growth(#Nr(o)) = limsupr→∞log#Nr(o)/logr, growth––––––––(#Nr(o)) = liminfr→∞log#Nr(o)/logr.

has finite polynomial growth if . If the upper and lower growth rates are equal, the common value is called the growth rate of .

For , one has and , where . This implies that and do not depend on the choice of the point .

In various situations in this paper, some weight in can be assigned to each point of . In these cases, it is natural to redefine the growth rate by considering the weights; i.e., by replacing with the sum of the weights of the points in . This will be formalized below using the notion of equivariant processes of Subsection I.2.6.

In the following definition, weights should be defined for all discrete spaces . However, if a random pointed discrete space is considered, it is enough to define weights in almost every realization (see Subsection I.2.6 for more on the matter). Also, given , the weights are allowed to be random.

###### Definition 2.1.

An equivariant weight function is an equivariant process (Definition I.2.6) with values in . For a discrete space and , the (random) value is called the weight of . Also, for , let

 w(S):=wD(S):=∑v∈Sw(v).

The last equation shows that one could also call an equivariant measure.

Assume is a unimodular discrete space (Subsection I.2.5). Lemma I.2.28 shows that is a random pointed marked discrete space and is unimodular. Also, one can let be undefined for a class of discrete spaces, as long as is almost surely defined.

In the following, the term ‘ is non-degenerate (i.e., not identical to zero) with positive probability’ means that

 P[∃v∈D:wD(v)≠0]>0. (2.1)

In the case when is unimodular, Lemma I.2.30 implies that the above condition is equivalent to

 E[w(o)]>0.

Also, the term ‘ is non-degenerate a.s.’ means that

 P[∃v∈D:wD(v)≠0]=1.

### 2.1 Unimodular Mass Distribution Principle

###### Theorem 2.2 (Mass Distribution Principle).

Let be a unimodular discrete space.

1. Let and assume there exists an equivariant weight function such that the weight of the ball with center and radius satisfies

 ∀r≥M:w(Nr(o))≤crα,a.s. (2.2)

Then, defined in (I.3.3) satisfies

 HαM(D)≥1cE[w(o)].
2. If in addition to (2.2), is non-degenerate with positive probability (see (2.1)), then

 udimH(D)≤α.
###### Proof.

Let be an arbitrary equivariant covering such that a.s. By the assumption on , a.s. Therefore,

 E[R(o)α]≥1cE[w(NR(o))]. (2.3)

By letting , one gets . Also, a.s., where the last inequality follows from the fact that is a covering. Therefore, the mass transport principle I.2.3 implies that (recall that by convention, is the empty set when ). So by (2.3), one gets . Since this holds for any , one gets that and the first claim is proved.

If, with positive probability, is non-degenerate, then Lemma I.2.30 implies that with positive probability. So . Therefore, and the second claim is proved. ∎

###### Example 2.3.

Theorem 2.2, applied to the counting measure , implies that and . As already proved in Propositions I.3.24 and I.3.38, equality holds in both.

###### Remark 2.4.

The assumption (2.2) in Theorem 2.2 implies a uniform polynomial bound on the size of all balls by the fact that what holds a.s. at the root, holds a.s. at all points (Lemma I.2.30).
In practice, this assumption is not as applicable as its continuum counterpart (Lemma 1.9 in [9]), except for some simple examples. The unimodular Billingsley lemma in the next subsection will actually be more useful. See also Lemma 2.5 below.

### 2.2 Unimodular Billingsley Lemma

The main result of this subsection is Theorem 2.8. It is based on Lemmas 2.5 and 2.6 below. Lemma 2.5 is a stronger version of the mass distribution principle (Theorem 2.2).

###### Lemma 2.5 (An Upper Bound).

Let be a unimodular discrete space.

1. If and is an equivariant weight function such that

 limsupr→∞w(Nr(o))rα≤c,a.s.,

then

2. In addition, if is non-degenerate with positive probability (see (2.1)), then .

###### Proof.

Let be arbitrary. The assumption implies that a.s. For , let

 Am:={v∈D:∀r≥m:w(Nr(v))≤c′rα},

which is an increasing sequence of equivariant subsets. Since ,

 limm→∞P[o∈Am]=1. (2.4)

Let be an equivariant covering such that a.s. One has

 E[R(o)α]≥E[R(o)α1{NR(o)∩Am≠∅}]. (2.5)

If , then and hence . In the next step, assume that this is the case. Let be an arbitrary point in . By the definition of , one gets that for all , Since , it follows that . Therefore, (2.5) gives

 E[R(o)α]≥12αc′E[w(NR(o))1{NR(o)∩Am≠∅}]. (2.6)

By letting , one gets that . Also, since there is a ball that covers a.s., one has a.s. Therefore, the mass transport principle (I.2.3) and (2.6) imply that

 E[R(o)α]≥12αc′E[w(o)1{o∈Am}].

This implies that . Using (2.4) and letting tend to infinity gives . Since is arbitrary, the first claim is proved.

Part (ii) of the claim is proved by arguments similar to those in Theorem 2.2. ∎

###### Lemma 2.6 (A Lower Bound).

Let be a unimodular discrete space, and . Let be an arbitrary equivariant weight function such that .

1. If a.s., then .

2. If a.s., then .

3. If , then .

4. If , then .

###### Proof.

The proofs of the first two parts are very similar. So the second part is proved first.

(ii). Let , and be such that . Fix . Let be the equivariant subset obtained by selecting each point with probability independently. Let if , if , and otherwise. Then is an equivariant covering. It is shown below that .

Let . By the assumption, a.s. One has

 E[Rn(o)γ] = nγP[o∈SD]+P[Nn(o)∩SD=∅] = nγE[1∧n−βw(o)]+E⎡⎣∏v∈Nn(o)(1−(1∧n−βw(v)))⎤⎦ ≤ nγ−βE[w(o)]+E[exp(−n−βw(Nn(o)))] = nγ−βE[w(o)]+E[exp(−n−βw(Nn(o)))|M

Therefore, . It follows that . Since is arbitrary, this implies .

(i). Only a small change is needed in the above proof. For , let if either or , and let otherwise. Note that is a covering by balls of equal radii. By the same computations and the assumption , one gets

 P[Rn(o)≠0]≤n−βE[w(o)]+exp(−nκ−β),

which is of order for large . This implies that . Since is arbitrary, one gets and the claim is proved.

(iii). Let be arbitrary. It will be proved below that there is a sequence such that . If so, by a slight modification of the proof of Part (ii), one can find a sequence of equivariant coverings such that and (iii) is proved.

Let be arbitrary. By the assumption, there is and such that . So

 E[exp(−r−βw(Nr(o)))] ≤ E[exp(−r−βw(Nr(o)))|w(Nr(o))>δrα] +P[w(Nr(o))≤δrα] ≤ exp(−δrα−β)+ϵ.

Note that for fixed and as above, can be arbitrarily large. Now, choose large enough for the right hand side to be at most . This shows that can be arbitrarily small and the claim is proved.

(iv). As before, let if either or , and let otherwise. The calculations in the proof of part (ii) show that

 P[Rn(o)≠0]≤n−βE[w(o)]+E[exp(−n−βw(Nn(o)))].

Now, the assumption implies the claim. ∎

###### Remark 2.7.

The assumption in part (iii) of Lemma 2.6 is equivalent to the condition that a subsequence of the family of random variables converges to zero in probability (as ). Also, from the proof of the lemma, one can see that this assumption is equivalent to

 liminfn→∞E[exp(−w(Nn(o))nα)]=0.
###### Theorem 2.8 (Unimodular Billingsley Lemma).

Let be a unimodular discrete metric space. For all equivariant weight functions such that

 0

one has

1. If the upper and lower growth rates of are almost surely constant (e.g., when is ergodic), then, almost surely,

 growth––––––––(w(Nr(o)))≤ udimH(D) ≤¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯growth(w(Nr(o))) ≤¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯growth(E[w(Nr(o))]).
2. In general,

 essinf(growth––––––––(w(Nr(o)))) ≤ udimH(D) ≤ essinf(¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯growth(w(Nr(o)))) ≤ ¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯growth(E[w(Nr(o))]).
###### Proof.

The first part is implied by the second one. So it is enough to prove the second part.

The first inequality is implied by part (ii) of Lemma 2.6. The last inequality is implied by Lemma A.1. For the second inequality, assume that with positive probability. On this event, one has for large ; i.e., a.s. Now, Lemma 2.5 implies that . This proves the result. ∎

###### Corollary 2.9.

Under the assumptions of Theorem 2.8, if exists and is constant a.s., then

 udimH(D)=growth(w(Nr(o))).

Subsection 3.5.1 below provides an example where .

###### Remark 2.10.

In fact, the assumption in Theorem 2.8 is only needed for the lower bound while the assumption is only needed for the upper bound. These assumptions are also necessary as shown below.
For example, assume is a point-stationary point process in (see Example I.2.18). For , let be the sum of the distances of to its next and previous points in . This equivariant weight function satisfies for all , and hence . But can be strictly less than 1 as shown in Subsection 3.3.1.
Also, the condition that is non-degenerate a.s. is trivially necessary for the upper bound.

###### Remark 2.11.

In many examples, it is enough to consider in Billingsley’s lemma (i.e., ). Examples where other weight functions are used are 2-ended trees (Subsection 3.1.1), point-stationary point processes (Proposition 2.17), and embedded spaces (Subsection 4.5).

###### Remark 2.12.

A converse to the unimodular Billingsley lemma is the unimodular Frostman lemma, which will be discussed in Section 4.

###### Remark 2.13.

Without the assumption of part (i) of Theorem 2.8, the claim is still valid for the sample Hausdorff dimension of , which will be discussed in Part III.

Here is a first application of the unimodular Billingsley lemma.

###### Proposition 2.14.

Let be a unimodular random graph equipped with the graph-distance metric. If is infinite almost surely, then and else, .

###### Proof.

If is infinite a.s., then for , one has for all . So part (i) of Lemma 2.6 implies the first claim.

For all discrete spaces , let if is finite and otherwise. One has . If is finite with positive probability, then . Therefore, the unimodular Billingsley lemma (Theorem 2.8) implies that , which in turn implies the second claim. ∎

###### Remark 2.15.

The upper bound in Theorem 2.8 is analogous to Billingsley’s lemma in the continuum setting (see e.g., Lemma 3.1 in [9]). It is interesting that there is no need to assume is embedded in or the bounded subcover property holds (see Remark 3.2 in [9]). Note also that does not depend on the origin in contrast to the analogous term in the continuum version.

### 2.3 Bounds for Point Processes

The next results use the following equivariant covering. Let be a discrete subset of equipped with the metric and . Let , be a point chosen uniformly at random in , and consider the partition of by cubes. Then, for each , choose a random element in independently (if the intersection is nonempty). The distribution of this random element should depend on the set in a translation-invariant way (e.g., choose with the uniform distribution or choose the least point in the lexicographic order). Let assign the value to the selected points and zero to the other points of . As in Example I.3.16, one can show that is an equivariant covering. Also, each point is covered at most times. So is -bounded (Definition I.3.12).

###### Theorem 2.16 (Minkowski Dimension in the Euclidean Case).

Let be a point-stationary point process in and assume the metric in is equivalent to the Euclidean metric. Then, for all equivariant weight functions such that a.s., one has

 ¯¯¯¯¯¯¯¯¯¯¯¯udimM(Φ) = ¯¯¯¯¯¯¯¯¯¯¯¯¯¯decay(E[w(0)/w(Cr+Ur)])≤¯¯¯¯¯¯¯¯¯¯¯¯¯¯decay(E[w(0)/w(Nr(0))]) ≤ ¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯growth(E[w(Nr(0))]), udim––––––M(Φ) = decay––––––(E[w(0)/w(Cr+Ur)])≤decay––––––(E[w(0)/w(Nr(0))]) ≤ growth––––––––(E[w(Nr(0))]),

where is a uniformly at random point in independent of and .

###### Proof.

By Theorem I.3.41, one may assume the metric on is the metric without loss of generality. Given any , consider the equivariant covering described above, but when choosing a random element of , choose point with probability (conditioned on ). One gets

 P[0∈R]=E[w(0)w(Cr+Ur)].

As mentioned above, is equivariant and uniformly bounded (for all ). So Lemma I.3.13 implies both equalities in the claim. The inequalities are implied by the facts that and

 E[w(0)w(Nr(0))]E[w(Nr(0))]≥E[√w(0)]2>0,

which is implied by the Cauchy-Schwartz inequality. ∎

In many examples, the case where is used. An example where the decay rate of is strictly smaller than the growth rate of can be found in Example III.5.2 for suitable parameters. However, this example is not ergodic (see Remark I.3.28).

###### Proposition 2.17.

If is a point-stationary point process in and the metric on is equivalent to the Euclidean metric, then .

###### Proof.

One may assume the metric on is the metric without loss of generality. Let and be a random point in chosen uniformly. For all discrete subsets and , let be the cube containing of the form (for ) and . Now, is an equivariant weight function. The construction readily implies that . Moreover, by , one has . Therefore, the unimodular Billingsley lemma (Theorem 2.8) implies that . ∎

###### Proposition 2.18.

If is a stationary point process in and is its Palm version, then

 udimM(Φ)=udimH(Φ)=k.

Moreover,

 Mk(Φ)=2kρ(Ψ),

where is the intensity of .

Notice that if , the claim is implied by Theorem I.3.48. The general case is treated below.

###### Proof.

For the first claim, by Proposition 2.17, it is enough to prove that . Let be a shifted square lattice independent of (i.e., , where is chosen uniformly, independently of ). Let . Since is a superposition of two independent stationary point processes, it is a stationary point process itself (see e.g., [10]). By letting , the Palm version of is obtained by the superposition of and an independent stationary lattice with probability (heads), and the superposition of and with probability (tails).

For all , there exists a disjoint -covering of (Example I.3.15). In both cases (heads or tails) above, one can consider this covering as a random subset of the shifted lattice. It is easy to see that it provides an equivariant -covering of (note that by enlarging the balls, all of is covered). Also, the probability of having a ball centered at the origin is . It follows that .

Note that has two natural equivariant subsets which, after conditioning to contain the origin, have the same distributions as and respectively. Therefore, one can use Theorem I.3.48 to deduce that . Therefore, Proposition 2.17 implies that .

Also, by using Theorem I.3.48 twice, one gets and . Therefore,

 Mk(Φ)=p1−pMk(Zk)=2kρ(Ψ),

where the last equality is by Proposition I.3.38. So the claim is proved. ∎

The last claim of Proposition 2.18 suggests the following, which is verified when in the next proposition.

###### Conjecture 2.19.

If is a point-stationary point process in which is not the Palm version of any stationary point process, then .

###### Proposition 2.20.

Conjecture 2.19 is true when .

###### Proof.

Denote as such that and for each . Then, the sequence is stationary under shifting the indices (see Example I.2.18). The assumption that is not the Palm version of a stationary point process is equivalent to (see [11] or Proposition 6 of [17]). Indeed, if , then one could bias the probability measure by (Definition I.B.1) and then shift the whole process by , where is chosen uniformly and independently.

Since , Birkhoff’s pointwise ergodic theorem [20] implies that . This in turn implies that . Therefore, Lemma 2.5 gives that ; i.e., . ∎

### 2.4 Connections to Birkhoff’s Pointwise Ergodic Theorem

This subsection discusses a corollary of the unimodular Billingsley lemma. The reader may skip it at first reading.

The following corollary of the unimodular Billingsley lemma is of independent interest. Note that the statement does not involve dimension.

###### Theorem 2.21.

Let be a unimodular discrete space. For any two equivariant weight functions and , if and is non-degenerate a.s., then

 growth––––––––(w1(Nr(o)))≤¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯growth(w2(Nr(o))),a.s.

In particular, if and have well defined growth rates, then their growth rates are equal.

Note that the condition is necessary as shown in Remark 2.10.

###### Proof.

Let be arbitrary and

 A:={[D,o]∈D∗:growth––––––––(w1(Nr(o)))>¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯growth(w2(Nr(o)))+ϵ}.

It can be seen that is a measurable subset of . Assume . Denote by the random pointed discrete space obtained by conditioning on . Since does not depend on the root (i.e., if , then ), by a direct verification of the mass transport principle (I.2.2), one can show that is unimodular. So by using the unimodular Billingsley lemma (Theorem 2.8) twice, one gets

 essinf(growth––––––––(w1(Nr(o′))))≤udimH(D′)≤essinf(¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯growth(w2(Nr(o′)))).

By the definition of , this contradicts the fact that a.s. So and the claim is proved. ∎

###### Remark 2.22.

Theorem 2.21 is a generalization of a weaker form of Birkhoff’s pointwise ergodic theorem as explained in the following.

1. If , then and are stationary under the shift (see Example I.2.18). Therefore, Birkhoff’s pointwise ergodic theorem (see e.g., [20]) implies that a.s. This equation is stronger than the claim of Theorem 2.21 in this case.

2. If is a point-stationary point process in , then the above argument still holds by using stationarity under shifting the origin to its next point (see Example I.2.18).

3. If is the Palm version of a stationary point process, the cross-ergodic theorem (see e.g., [4]) also gives the property stated in (i).

Note that amenability is not assumed in Theorem 2.21. But the claim is weaker since nothing can be said about .

## 3 Examples

The structure of this section is analogous to that of Section I.4. It provides further results on the examples introduced there. Some new examples are also presented.

### 3.1 General Unimodular Trees

#### 3.1.1 Unimodular Two-Ended Trees

Here, another proof is given for the fact that the Minkowski and Hausdorff dimensions of any unimodular two-ended tree are equal to one (Theorem I.4.2) using the unimodular Billingsley lemma.

Let be a unimodular two-ended tree. For all two-ended trees and , let be 1 if belongs to the trunk of and 0 otherwise. It can be seen that is an equivariant process (Definition I.2.21). Let be the distance of to the trunk of . For larger than , one has . Therefore, the unimodular Billingsley lemma (Theorem 2.8) implies that . On the other hand, Proposition 2.14 implies that . Therefore, .

#### 3.1.2 Unimodular Trees with Infinitely Many Ends

The following conjecture is of independent interest beyond its connections to the dimension. The authors believe it is new.

###### Conjecture 3.1.
1. There is no unimodular random tree with polynomial growth and infinitely many ends a.s.

2. There is no equivariant metric (Definition I.3.40) on the -regular tree (for ) that has polynomial growth.

By the unimodular Billingsley lemma (Theorem 2.8), this conjecture is implied by Conjectures I.4.7 and I.4.8.

A stronger form of part (i) of the above conjecture is that there is no unimodular random tree with infinitely many ends a.s. such that .

The following example and proposition provide special cases where Conjecture I.4.8 is known to be true. Another example is provided in Subsection 3.2.1 below.

In the following, the ball of radius under the metric and centered at is denoted by .

###### Example 3.2.

By assigning i.i.d. random lengths to the edges on the -regular tree (as in Example I.2.24), one gets an equivariant metric on , denoted by . The claim is that for , one has , which implies that Conjecture I.4.8 holds in this case.
By Theorem I.3.41, one can replace the distribution of the lengths with another distribution which is stochastically larger. So one can assume the distribution of the lengths is non-lattice without loss of generality. Therefore, Theorem 21.1 of [14] for supercritical age-dependent branching processes implies that the limit of as exists a.s. and is positive for some . Therefore, Lemma 2.6 implies that and the claim is proved.

Recall from Subsection I.4.1.4 that an equivariant metric is said to be generated by equivariant edge lengths (or a geodesic metric) if for every path , one has .

###### Proposition 3.3.

Let be an equivariant metric on the 3-regular tree which is generated by equivariant edge lengths. If the random variable has finite moment of order , then . In particular, if it has finite moments of any order, then .

###### Proof.

It is enough to assume for all since increasing the edge lengths does not increase the dimension (Theorem I.3.41). Consider the following equivariant weight function on :

 w(u):=C∑v∼ud′(u,v)α, (3.1)

where is a constant such that

 ∀x∈[0,1]:Cxα+(1−x)α≥12. (3.2)

It is easy to see that such a exists. Now, Lemma 3.4 below, which is a deterministic result, implies that a.s. for every . Also, the assumption on implies that has finite mean. So Lemma 2.6 implies that and the claim is proved. ∎

The following lemma is used in the last proof.

###### Lemma 3.4.

Let and be a deterministic rooted tree such that and for all . Let be a metric on which is generated by a function on the edges such that . Define the weight function on and the constant by (3.1) and (3.2) respectively. Then, for all , one has .

###### Proof.

For , let be the infimum value of for all trees with the stated conditions. So one should prove . The claim is true for . Also, if , one has and the claim is trivial. The proof uses induction on . Assume that and for all , one has . For , let be the connected component containing when the edge is removed. It can be seen that