The Annals of Applied Probability

2016, Vol. 26, No. 3, 1807–1836

DOI: 10.1214/15-AAP1134
© Institute of Mathematical Statistics, 2016

HACK’S LAW IN A DRAINAGE NETWORK MODEL:

A BROWNIAN WEB APPROACH

B Y R AHUL ROY , K UMARJIT S AHA AND A NISH S ARKAR

Indian Statistical Institute
Hack [Studies of longitudinal stream profiles in Virginia and Maryland
(1957). Report], while studying the drainage system in the Shenandoah valley
and the adjacent mountains of Virginia, observed a power law relation l ∼
a 0.6 between the length l of a stream from its source to a divide and the
area a of the basin that collects the precipitation contributing to the stream
as tributaries. We study the tributary structure of Howard’s drainage network
model of headward growth and branching studied by Gangopadhyay, Roy
and Sarkar [Ann. Appl. Probab. 14 (2004) 1242–1266]. We show that the
exponent of Hack’s law is 2/3 for Howard’s model. Our study is based on a
scaling of the process whereby the limit of the watershed area of a stream is
area of a Brownian excursion process. To obtain this, we define a dual of the
model and show that under diffusive scaling, both the original network and
its dual converge jointly to the standard Brownian web and its dual.

1. Introduction. River basin geomorphology is a very old subject of study

initiated by Horton [14]. Hack [13], studying the drainage system in the Shenan-
doah valley and the adjacent mountains of Virginia, observed a power law relation
(1) l ∼ a 0.6
between the length l of a stream from its source to a divide and the area of the
basin a that collects the precipitation contributing to the stream as tributaries. Hack
also corroborated this power law through the data gathered by Langbein [17] of
nearly 400 different streams in the northeastern United States. This empirical re-
lation (1) is widely accepted nowadays albeit with a different exponent (see Gray
[12], Muller [22]) and is called Hack’s law. Mandelbrot [21] mentions Hack’s law
to strengthen his contention that “if all rivers as well as their basins are mutually
similar, the fractal length-area argument predicts (river’s length)1/D is propor-
tional to (basin’s area)1/2 ” where D > 1 is the fractal dimension of the river. In
this connection, it is worth remarking that the Hurst exponent in fractional Brow-
nian motion and in time series analysis arose from the study of the Nile basin by
Hurst [15] where he proposed the relation l⊥ = l0.9 as that governing the width,
l⊥ , and the length, l , of the smallest rectangular region containing the drainage
system.

Received February 2015; revised July 2015.

MSC2010 subject classifications. 60D05.
Key words and phrases. Brownian excursion, Brownian meander, Brownian web, Hack’s law.
1807
1808 R. ROY, K. SAHA AND A. SARKAR

Various statistical models of drainage networks have been proposed (see

Rodriguez-Iturbe and Rinaldo [26] for a detailed survey). In this paper, we study
the tributary structure of a two-dimensional drainage network called the Howard’s
model of headward growth and branching (see Rodriguez-Iturbe and Rinaldo [26]).
Our study is based on a scaling of the process and we obtain the watershed area of
a stream as the area of a Brownian excursion process. This gives a statistical expla-
nation of Hack’s law and justifies the remark of Giacometti et al. [11]: “From the
results, we suggest that a statistical framework referring to the scaling invariance
of the entire basin structure should be used in the interpretation of Hack’s law.”
We first present an informal description of the model: suppose that the vertices
of the d-dimensional lattice Zd are open or closed with probability p (0 < p < 1)
and 1 − p, respectively, independently of all other vertices. Each open vertex u ∈
Zd represents a water source and connects to a unique open vertex v ∈ Zd . These
edges represent the channels through which water can flow. The connecting vertex
v is chosen so that the dth coordinate of v is one more than that of u and v has
the minimum L1 distance from u. In case of nonuniqueness of such a vertex, we
choose one of the closest open vertices with equal probability, independently of
everything else.
Let V denote the set of open vertices and h(u) denote the uniquely chosen ver-
tex to which u connects, as described above. Set u, h(u) as the edge (channel)
connecting u and h(u). From the construction, it follows that the random graph,
G = (V , E) with edge set E := {u, h(u) : u ∈ V }, does not contain any circuit.
This model has been studied by Gangopadhyay, Roy and Sarkar [10] and the fol-
lowing results were obtained.

T HEOREM 1.1. Let 0 < p < 1.

(i) For d = 2 and d = 3, G consists of one single tree almost surely, and for
d ≥ 4, G is a forest consisting of infinitely many disjoint trees almost surely.
(ii) For any d ≥ 2, the graph G contains no bi-infinite path almost surely.

In this paper, we consider only d = 2. Before proceeding further, we present

a formal description for d = 2 which will be used later. Fix 0 < p < 1 and let
{Bu : u = (u(1), u(2)) ∈ Z2 } be an i.i.d. collection of Bernoulli random vari-
ables with success probability p. Set V = {u ∈ Z2 : Bu = 1}. Let {Uu : u ∈ Z2 }
be another i.i.d. collection of random variables, independent of the collection of
random variables {Bu : u ∈ Z2 }, taking values in the set {1, −1}, with P(Uu =
1) = P(Uu = −1) = 1/2. For a vertex (x, t) ∈ Z2 , we consider k0 = min{|k| : k ∈
Z, B(x+k,t+1) = 1}. Clearly, k0 is almost surely finite. Now, we define
⎧
⎨ (x + k0 , t + 1) ∈ V , if (x − k0 , t + 1) ∈
/ V,
h(x, t) := (x − k0 , t + 1) ∈ V , if (x + k0 , t + 1) ∈
/ V,
⎩
(x + U(x,t) k0 , t + 1) ∈ V , if (x ± k0 , t + 1) ∈ V .
 HACK’S LAW IN A DRAINAGE NETWORK MODEL 1809

For any k ≥ 0, let

 
hk+1 (x, t) := h hk (x, t) with h0 (x, t) := (x, t),

Ck (x, t) := (y, t − k) ∈ V : h (y, t − k) = (x, t) , if (x, t) ∈ V ,
k

∅, otherwise,

C(x, t) := Ck (x, t).

k≥0

Here, hk (x, t) represents the “kth generation progeny” of (x, t), the sets Ck (x, t)
and C(x, t) denote, respectively, the set of kth generation ancestors and the set of
all ancestors of (x, t); C(x, t) = ∅ if (x, t) ∈/ V . In the terminology of drainage
network, C(x, t) represents the region of precipitation, the water from which is
channelled through the open point (x, t) (see Figure 1). From Theorem 1.1(ii), we
have that C(x, t) is finite almost surely.
Now, we define

L(x, t) := inf k ≥ 0 : Ck (x, t) = ∅ ,

as the “length of the channel,” which as earlier is finite almost surely. We observe
that for any (x, t) ∈ Z × Z, L(x, t) ≥ 0 and the distribution of L(x, t) does not
depend upon (x, t). Our first result is about the length of the channel. We remark
here that Newman, Ravishankar and Sun [23] has a similar result in a set-up which
allows crossing of paths.

F IG . 1. The bold vertices on the line y = t − 3 constitute the set C3 (x, t) and all the bold vertices
together constitute the cluster C(x, t).
1810 R. ROY, K. SAHA AND A. SARKAR

T HEOREM 1.2. We have

√   1
lim nP L(0, 0) > n = √ ,
n→∞ γ0 π
(1−p)(2−2p+p 2 )
where γ02 := γ02 (p) = p 2 (2−p)2
.

Next, we define
 
max u : (u, t − k) ∈ Ck (x, t) , if 0 ≤ k < L(x, t), (x, t) ∈ V ,
rk (x, t) :=
0, otherwise,
 
min u : (u, t − k) ∈ Ck (x, t) , if 0 ≤ k < L(x, t), (x, t) ∈ V ,
lk (x, t) :=
0, otherwise,
Dk (x, t) := rk (x, t) − lk (x, t).
The quantity Dk (x, t) denotes the width of the set of all kth generation ancestors
of (x, t). We define the width process Dn(x,t) (s) and the cluster process Kn(x,t) (s)
for s ≥ 0 as follows: for k = 0, 1, . . . and k/n ≤ s ≤ (k + 1)/n,
Dk (x, t) (ns − [ns])  
Dn(x,t) (s) := √ + √ Dk+1 (x, t) − Dk (x, t) ,
γ0 n γ0 n
(2)
#Ck (x, t) (ns − [ns])  
Kn(x,t) (s) := √ + √ #Ck+1 (x, t) − #Ck (x, t) ,
γ0 n γ0 n
(x,t)
where γ0 > 0 is as in the statement of Theorem 1.2. In other words, Dn (s) [resp.,
(x,t) √ √
Kn (s)] is defined Dk (x, t)/(γ0 n) [resp., #Ck (x, t)/(γ0 n)] at time points
s = k/n and, at other time points defined by linear interpolation. The distributions
of both Dn(x,t) and Kn(x,t) are independent of (x, t).
To describe our results, we need to introduce two processes, Brownian me-
ander and Brownian excursion, studied by Durrett, Iglehart and Miller [7]. Let
{W (s) : s ≥ 0} be a standard Brownian motion with W (0) = 0. Let τ1 := sup{s ≤
1 : W (s) = 0} and τ2 := inf{s ≥ 1 : W (s) = 0}. Note that τ1 < 1 and τ2 > 1 al-
most surely. The standard Brownian meander, W + (s), and the standard Brownian
excursion, W0+ (s), are given by
|W (τ1 + s(1 − τ1 ))|
(3) W + (s) := √ , s ∈ [0, 1],
1 − τ1
|W (τ1 + s(τ2 − τ1 ))|
(4) W0+ (s) := √ , s ∈ [0, 1].
τ2 − τ1
Both of these processes are a continuous nonhomogeneous Markov process (see
Durrett and Iglehart [6] and references therein). Further, W + (0) = 0 and, for x ≥ 0,
P(W + (1) ≤ x) = 1 − exp(−x 2 /2), that is, W + (1) follows a Rayleigh distribution.
 HACK’S LAW IN A DRAINAGE NETWORK MODEL 1811

We also need some random variables obtained as functionals of these two pro-
cesses. In particular, let
 1

I0+ := W0+ (t) dt and M0+ := max W0+ (t) : t ∈ [0, 1] .
0
Louchard and Janson [20] showed that, as x → ∞, the distribution function and
the density are, respectively, given by
√ √
 +  6 6  2 72 6 2  
P I0 > x ∼ √ x exp −6x and fI + (x) ∼ √ x exp −6x 2 .
π 0 π
The random variable M0+ is continuous, having a strictly positive density on (0, ∞)
(see Durrett and Iglehart [6]) and for x > 0,
  ∞
   
2  
P M0+ ≤x =1+2 exp −(2kx) /2 1 − (2kx)
2
with E M0+ = π/2.
k=1
For f ∈ C[0, ∞), let f |[0,1] denotes the restriction of f over [0, 1]. Our next
result is about the weak convergence of the width process Dn(0,0) |[0,1] and the clus-
(0,0)
ter process Kn |[0,1] under diffusive scaling. Here and subsequently, as is com-
monly used in statistics, we use the notation X|Y to denote the conditional random
variable X given Y .

T HEOREM 1.3. As n → ∞, we have:

√
(i) Dn |[0,1] |1{L(0,0)>n} ⇒ 2W + ,
(0,0)
P
(ii) sup{|pDn(0,0) (s) − Kn(0,0) (s)| : s ∈ [0, 1]}|1{L(0,0)>n} → 0.

The following corollary is an immediate consequence of Theorem 1.3.

C OROLLARY 1.3.1. For u > 0, as n → ∞ we have:

√ √
(i) nP(#Cn (0, 0) > nγ0 u) → γ √ 1
exp(−u2 /4p2 ),
n 0 π √
(ii) P( k=0 #Ck (0, 0) > n3/2 γ0 u|L(0, 0) > n) → P(p 2I + > u).

Before we proceed to state Theorem 1.4, we recall some results regarding

random vectors whose distribution functions have regularly varying tails (see
Resnick [25], page 172). A random vector Z on (0, ∞)d with a distribution func-
tion F has a regularly varying tail if, as n → ∞, there exists a sequence bn → ∞
v
such that nP{Z/bn ∈ ·} → ν(·) for some ν ∈ M+ where M+ := {μ : μ is a non-
v
negative Radon measure on (0, ∞)d }. Here, → denotes vague convergence. It is
in this context that Theorem 1.4 obtains a regularly varying tail for the distribution
of (L(x, t), (#C(x, t))2/3 ); which justifies that the exponent of Hack’s law is 2/3
for Howard’s model. In addition, we obtain a scaling law, with a Hack exponent
1812 R. ROY, K. SAHA AND A. SARKAR

of 1/2, for the length of the stream, vis-à-vis the maximum width of the region of
precipitation, that is,

(5) Dmax (0, 0) := max Dk (0, 0) : 0 ≤ k < L(0, 0) .
It should be noted that Leopold and Langbein [18] obtained an exponent of 0.64
through computer simulations.

T HEOREM 1.4. Let E := [0, ∞) × [0, ∞) \ {(0, 0)}. There exist measures μ
and ν on the Borel σ -algebra on E such that for any Borel set B ⊆ E we have
 
√ (L(0, 0), (#C(0, 0))2/3 )
(6) nP ∈ B → μ(B),
n
 
√ (L(0, 0), (Dmax (0, 0))1/2 )
(7) nP ∈ B → ν(B),
n
with μ and ν being given by
 √  
3 v v 3/2
μ(B) = √ fI + √ dv dt,
B 4 2πγ02 pt 3 0 γ0 p 2t 3
  
v v2
ν(B) = √ fM + √ dv dt
B 2π γ02 pt 2 0 γ0 p 2t
where fI + and fM + denote the density functions of I0+ and M0+ , respectively.
0 0
Moreover, for λ, τ > 0, we have
 
√ (L(0, 0), (#C(0, 0))α )
nP ∈ (τ, ∞) × (λ, ∞)
n
(8) ⎧
⎪
⎪ 2
⎪
⎪ 0, if α < ,
⎨ 3
= 1 2
⎪
⎪ 
⎪
⎪ , if α >
⎩ πτ γ 2 3
0
and
 
√ (L(0, 0), (Dmax (0, 0))α )
nP ∈ (τ, ∞) × (λ, ∞)
n
(9) ⎧
⎪
⎪ 0, 1
⎪
⎪ if α < ,
⎨ 2
= 1 1
⎪
⎪  , if α > .
⎪
⎪
⎩ πτ γ 2 2
0

The estimates of the densities fI + and fM + imply that μ and ν are finite mea-
0 0
sures on E. An immediate consequence of the above theorem is the following.
 HACK’S LAW IN A DRAINAGE NETWORK MODEL 1813

C OROLLARY 1.4.1. As n → ∞ for u > 0, we have:

√ √ 
(i) nP(#C(0, 0) > 2n3 γ0 pu) → 2√1π γ 0∞ t −3/2 F
 + (ut −3/2 ) dt,
I0
√ √ 0
 ∞ −3/2
(ii) nP(Dmax (0, 0) > 2nγ0 pu) → 2 π γ 0 t
√ 1 
FM + (ut −1/2 ) dt,
0 0

where FI + and FM + are the distribution functions of I0+ and M0+ , respectively,
0 0
 + := 1 − F + , F
and F  + := 1 − F + .
I0 I0 M0 M0

The proofs of the above theorems are based on a scaling of the process. In the
next section, we define a dual graph and show that as processes, under a suitable
scaling, the original and the dual processes converge jointly to the Brownian web
and its dual in distribution (the double Brownian web). This invariance principle
is used in Sections 3 and 4 to prove the theorems. In this connection, it is worth
noting that in Proposition 2.7, we have provided an alternate characterization of
the dual of Brownian web which is of independent interest. This characterization
is suitable for proving the joint convergence of coalescing noncrossing path family
and its dual to the double Brownian web and has been used in Theorem 2.9 to
achieve the required convergence.
We should mention here that the Brownian web appears as a universal scaling
limit for various network models (see Fontes et al. [9], Ferrari, Fontes and Wu [8],
Coletti, Fontes and Dias [5]). It is reasonable to expect that with suitable modifi-
cations our method will give similar results in other network models. Our results
will hold for any network model which admits a dual and satisfies (i) conditions
listed in Remark 2.1, (ii) the scaled model and its dual converges weakly to the
double Brownian web (see Section 2) and (iii) a certain sequence of counting ran-
dom variables are uniformly integrable (see Lemma 3.3). In this sense, our result
can be considered as a universality class result.

2. Dual process and the double Brownian web.

2.1. Dual process. For the graph G , we now describe a dual process such that
the set of ancestors C(x, t) (defined in the previous section) of a vertex (x, t) ∈ V
is bounded by two dual paths. The dependency inherent in the graph G implies
that, although the cluster is bounded by two dual paths, these paths are not given by
independent random walks. The dual vertices are precisely the mid-points between
two consecutive open vertices on each horizontal line {y = n}, n ∈ Z with each
dual vertex having a unique offspring dual vertex in the negative direction of the
y-axis. Before giving a formal definition, we direct the attention of the reader to
Figure 2.
For u ∈ Z2 , we define
  
Ju+ := inf k : k ≥ 1, u(1) + k, u(2) ∈ V ,
(10)   
Ju− := inf k : k ≥ 1, u(1) − k, u(2) ∈ V .
1814 R. ROY, K. SAHA AND A. SARKAR

F IG . 2. The black points are open vertices, the gray points are the vertices of the dual process and
the gray (dashed) paths are the dual paths.

Next, we define r(u) := (u(1) + Ju+ , u(2)) and l(u) := (u(1) − Ju− , u(2)), as the
first open point to the right (open right neighbour) and the first open point to
the left (open left neighbour) of u, respectively. For (x, t) ∈ V , let r̂(x, t) :=
+
(x + J(x,t) ˆ t) := (x − J − /2, t) denote, respectively, the right dual
/2, t) and l(x, (x,t)
neighbour and the left dual neighbour of (x, t) in the dual vertex set. Finally, the
dual vertex set is given by

ˆ t) : (x, t) ∈ V .
V := r̂(x, t), l(x,
For a vertex (u, s) ∈ V, let (v, s − 1) ∈ V be such that the straight line segment
joining (u, s) and (v, s − 1) does not cross any edge in G . The dual edges are edges
joining all such (u, s) and (v, s − 1). Formally, for (u, s) ∈ V, we define

a l (u, s) := sup z : (z, s − 1) ∈ V , h(z, s − 1)(1) < u ,
(11) 
a r (u, s) := inf z : (z, s − 1) ∈ V , h(z, s − 1)(1) > u

and set ĥ(u, s) := ((a l (u, s) + a r (u, s))/2, s − 1). Note that (a r (u, s), s − 1) and
(a l (u, s), s − 1) are the nearest vertices in V to the right and left, respectively, of
the dual vertex ĥ(u, s). Finally, the edge set of the dual graph G := (V, E)  is given
by
 
 := (u, s), ĥ(u, s) : (u, s) ∈ V
E  .

R EMARK 2.1. Note that the vertex set of the dual graph is a subset of 12 Z × Z.
Before we proceed, we list some properties of the graph G and its dual G.
(1) G uniquely specifies the dual graph G and the dual edges do not intersect
the original edges. The construction ensures that G does not contain any circuit.
 HACK’S LAW IN A DRAINAGE NETWORK MODEL 1815

(2) For (x, t) ∈ V , the cluster C(x, t) is enclosed within the dual paths starting
ˆ t). The boundedness of C(x, t) for every (x, t) ∈ V implies
from r̂(x, t) and l(x,
that these two dual paths coalesce, thus G is a single tree.
(3) Since paths starting from any two open vertices in the original graph coa-
lesce and the dual edges do not cross the original edges, there is no bi-infinite path
in G.

We now obtain a Markov process from the dual paths. Fix (u, s) ∈ V and for
k ≥ 1, set ĥk (u, s) := ĥ(ĥk−1 (u, s)) where ĥ0 (u, s) := (u, s). Letting X(u,s) denote
k
the first coordinate of ĥk (u, s), it may be observed that X  (u,s)
is a function (u,s)
of X
k+1 k
and the collection of random variables {(Bu , Uu ) : u(2) = s − k − 1 ∈ Z}. Thus, by
the random mapping representation (see, e.g., Levin, Peres, and Wilmer [19]) we
have the following.

P ROPOSITION 2.2. For (u, s) ∈ V, the process {X

(u,s) : k ≥ 0} is a time ho-
k
mogeneous Markov process.

Before we proceed, we make the following observations about the transition

probabilities of the Markov process. Let G be a geometric random variable taking
values in {1, 2, . . .}, that is, P(G = l) = p(1 − p)l−1 for l ≥ 1. For any u ∈ Z × Z,
the random variables Ju+ and Ju− are i.i.d. copies of the geometric random variable
G independent of Bu . Further, if u1 , u2 ∈ Z2 are such that u1 (1) ≥ u2 (1) − 1 and
u1 (2) = u2 (2), the random variables Ju+1 and Ju−2 are also independent. Now, for
u∈/ Z and, v ∈ Z/2, we have
   
(u,s) − X
PX (u,s) = u = P J +
(u,s) = v|X −
1 0 0 (u−1/2,s−1) − J(u+1/2,s−1) = 2v
(12)
= P(G1 − G2 = 2v),

where G1 and G2 are i.i.d. copies of G, defined above. If u ∈ Z and v ∈ Z/2, we

have, using notations from above
 (u,s) 
PX (u,s) = v|X
−X (u,s) = u
1 0 0
(13)
= (1 − p)P(G1 − G2 = 2v) + pP(G = 2v)/2 + pP(G = −2v)/2,

where G1 and G2 are as above. It is therefore obvious that the transition probabil-
(u,s) depend on whether the present state is an integer or not.
ities of X k
From equations (12) and (13), we state the following.

P ROPOSITION 2.3. For any (u, s) ∈ V, {X (u,s) : k ≥ 0} is an L2 -martingale

k
with respect to the filtration Fk := σ ({Bu , Uu : u ∈ Z2 , u(2) ≥ s − k}).
1816 R. ROY, K. SAHA AND A. SARKAR

2.2. Dual Brownian web. In this section, we briefly describe the dual Brown-
ian web W associated with W and present an alternate characterization of the dual
Brownian web W .
The Brownian web (studied extensively by Arratia [1, 2], Tóth and Werner [30],
Fontes et al. [9]) may be viewed as a collection of one-dimensional coalescing
Brownian motions starting from every point in the space time plane R2 . We recall
relevant details from Fontes et al. [9].
Let R2c denote the completion of the space time plane R2 with respect to the
metric
 
     tanh(x1 ) tanh(x2 ) 
ρ (x1 , t1 ), (x2 , t2 ) := tanh(t1 ) − tanh(t2 ) ∨  − .
1 + |t1 | 1 + |t2 | 
As a topological space R2c can be identified with the continuous image of
[−∞, ∞]2 under a map that identifies the line [−∞, ∞] × {∞} with the point
(∗, ∞), and the line [−∞, ∞] × {−∞} with the point (∗, −∞). A path π in R2c
with starting time σπ ∈ [−∞, ∞] is a mapping π : [σπ , ∞] → [−∞, ∞] ∪ {∗}
such that π(∞) = ∗ and, when σπ = −∞, π(−∞) = ∗. Also t → (π(t), t) is a
continuous map from [σπ , ∞] to (R2c , ρ). We then define
to be the space of all
paths in R2c with all possible starting times in [−∞, ∞]. The following metric, for
π1 , π2 ∈


 
d
(π1 , π2 ) := max tanh(σπ1 ) − tanh(σπ2 ),
 
 tanh(π1 (t ∨ σπ1 )) tanh(π2 (t ∨ σπ2 )) 
sup   −
1 + |t| 1 + |t| 
t≥σπ1 ∧σπ2

makes
a complete, separable metric space.

R EMARK 2.4. Convergence in this metric can be described as locally uniform

convergence of paths as well as convergence of starting times. Therefore, for any
ε > 0 and m > 0, we can choose ε1 (= f (ε, m)) > 0 such that for π1 , π2 ∈

with {(πi (t), t) : t ∈ [σπi , m]} ⊆ [−m, m] × [−m, m] for i = 1, 2, d

(π1 , π2 ) < ε1
implies that (π1 (σπ1 ), σπ1 ) − (π2 (σπ2 ), σπ2 )2 < ε and sup{|π1 (t) − π2 (t)| : t ∈
[max{σπ1 , σπ2 }, m]} < ε. We will use this later several times.

Let H be the space of compact subsets of (

, d
) equipped with the Haus-
dorff metric dH . The Brownian web W is a random variable taking values in the
complete separable metric space (H, dH ).
Before introducing the dual Brownian web, we require a similar metric space
on the collection of backward paths. As in the definition of
, let
 be the col-
lection of all paths π̂ with starting time σπ̂ ∈ [−∞, ∞] such that π̂ : [−∞, σπ̂ ] →
[−∞, ∞] ∪ {∗} with π̂ (−∞) = ∗ and, when σπ̂ = +∞, π̂(∞) = ∗. As earlier
 HACK’S LAW IN A DRAINAGE NETWORK MODEL 1817

 with
t → (π̂(t), t) is a continuous map from [−∞, σπ̂ ] to (R2c , ρ). We equip

the metric

 
 (π̂1 , π̂2 ) := max tanh(σπ̂1 ) − tanh(σπ̂2 ),
d

 
 tanh(π̂1 (t ∧ σπ̂1 )) tanh(π̂2 (t ∧ σπ̂2 )) 
sup   −
1 + |t| 1 + |t| 
t≤σπ̂1 ∨σπ̂2

 d  ) a complete, separable metric space. The complete separable met-

making (
,

ric space of compact sets of paths of

 is denoted by (H, d  ), where d  is the
H H
Hausdorff metric on H, and let B  be the corresponding Borel σ field.
H

). The Brownian web and its dual (W , W

2.3. Properties of (W , W ) is a (H ×
H, BH × B  ) valued random variable such that W and W uniquely determine each
H

other almost surely with W being equally distributed as −W , the Brownian web
rotated 180o about the origin. The interaction between the paths in W and W  is
that of Skorohod reflection (see Soucaliuc, Tóth and Werner [28]).
We introduce some notation to study the sets {π(t + s) : π ∈ W , σπ ≤ t} and
, σ ≥ t}. For a (H, BH ) valued random variable K and t ∈ R, let
{π̂(t −s) : π̂ ∈ W π̂
K := {π : π ∈ K and σπ ≤ t}. Similarly, for a (H
t− , B  ) valued random variable
H
  
K and t ∈ R, let K := {π̂ : π̂ ∈ K and σπ̂ ≥ t}. For t1 , t2 ∈ R, t2 > t1 and a
t+

(H, BH ) valued random variable K, define


MK (t1 , t2 ) := π(t2 ) : π ∈ K t1 − , π(t2 ) ∈ [0, 1] ;
(14)
ξK (t1 , t2 ) := #MK (t1 , t2 ),
that is, ξK (t1 , t2 ) denotes the number of distinct points in [0, 1] × t2 which are on
some path in K t1 − . We note that for t > 0, MW (t0 , t0 + t) = NW (t0 , t; 0, 1) as
defined in Sun and Swart [29]. It is known that for all t > 0 the random variable
ξW (t0 , t0 + t) is finite almost surely (see (E1 ) in Theorem 1.3 in Sun and Swart
[29]) with
  1
(15) E ξW (t0 , t0 + t) = √ .
πt
Moreover, from the known properties of (W , W ) the proof of the following propo-
sition is straightforward (for details, see Roy, Saha and Sarkar [27]).

P ROPOSITION 2.5. For any t0 < t1 , almost surely we have:

(i) MW (t0 , t1 ) ∩ Q = ∅;
(ii) each point in MW (t0 , t1 ) is of type (1, 1);
(iii) for each x ∈ MW (t0 , t1 ), there exists π1 , π2 ∈ W with σπ1 < t0 , σπ2 > t0
and π1 (t1 ) = π2 (t1 ) = x;
1818 R. ROY, K. SAHA AND A. SARKAR

(x,t1 ) (x,t1 )
(iv) for each x ∈ MW (t0 , t1 ), there exist exactly two paths π̂r and π̂l

in W starting from (x, t1 ) with π̂r(x,t1 ) (t) > π̂l(x,t1 ) (t) for all [t0 , t1 ).

There are several ways to construct W  from W . In this paper, we follow the
wedge characterization provided by Sun and Swart [29]. For π r , π l ∈ W with co-
r l
alescing time t π ,π and π r (max{σπ r , σπ l }) > π l (max{σπ r , σπ l }), the wedge with
right boundary π r and left boundary π l , is an open set in R2 given by
 
A = A πr, πl
(16)  r ,π l
:= (y, s) : max{σπ l , σπ r } < s < t π , π l (s) < y < π r (s) .
 is said to enter the wedge A from outside if there exist t1 and t2 with
A path π̂ ∈
,
σπ̂ > t1 > t2 such that (π̂(t1 ), t1 ) ∈
/A and (π̂(t2 ), t2 ) ∈ A.
From Theorem 1.9 in Sun and Swart [29], it follows that the dual Brownian web
 associated with the Brownian web W satisfies the following wedge characteri-
W
zation.

T HEOREM 2.6. ) be a Brownian web and its dual. Then almost

Let (W , W
surely
 = {π̂ : π̂ ∈

W  and does not enter any wedge in W from outside}.

Because of Theorem 2.6, for a (H × H , BH × B  ) valued random variable

H

(W , Z ) to show that Z = W , it suffices to check that Z satisfies the wedge condi-
tion. Here we present an alternate condition which is easier to check.

P ROPOSITION 2.7. , BH × B  ) valued random vari-

Let (W , Z ) be a (H × H H
able such that:
(1) for any deterministic (x, t) ∈ R2 , there exists a path π̂ (x,t) ∈ Z starting at
(x, t) and going backward in time almost surely;
(2) paths in Z do not cross paths in W almost surely, that is, there does not exist
any π ∈ W , π̂ ∈ Z and t1 , t2 ∈ (σπ , σπ̂ ) such that (π̂(t1 ) − π(t1 ))(π̂ (t2 ) − π(t2 )) <
0 almost surely;
(3) paths in Z and paths in W do not coincide over any time interval almost
surely, that is, for any π ∈ W and π̂ ∈ Z and for no pair of points t1 < t2 with
σπ ≤ t1 < t2 ≤ σπ̂ we have π̂(t) = π(t) for all t ∈ [t1 , t2 ] almost surely.
 almost surely.
Then Z = W

P ROOF. From conditions (2) and (3), we have that π̂ ∈ Z does not enter any
wedge in W from outside. Hence, Z ⊆ W . The argument for W  ⊆ Z follows from
the fish-trap technique introduced in the proof of Lemma 4.7 of Sun and Swart
⊆Z
[29]. It shows that W  almost surely for any (H, BH ) valued random variable
 HACK’S LAW IN A DRAINAGE NETWORK MODEL 1819

Z satisfying (i) paths in Z do not cross paths is W and (ii) for any deterministic
countable dense set, there exist paths in Z starting from every point of that dense
set (for details, see Roy, Saha and Sarkar [27]). 

2.4. Convergence to the double Brownian web. For any (x, t) ∈ V , the path
π (x,t) in the random graph G is obtained as the piecewise linear function π (x,t) :
[t, ∞) → R with π (x,t) (t + k) = hk (x, t)(1) for every k ≥ 0 and π (x,t) being linear
in the interval [t + k, t + k + 1]. Similarly, for (x, t) ∈ V, the dual path π̂ (x,t) is the
piecewise linear function π̂ (x,t) : (−∞, t] → R with π̂ (x,t) (t − k) = ĥk (x, t)(1)
for every k ≥ 0 and π̂ (x,t) being linear in the interval [t − k − 1, t − k]. Let X :=
{π (x,t) : (x, t) ∈ V } and X := {π̂ (x,t) : (x, t) ∈ V} be the collection of all possible
paths and dual paths admitted by G and G.
For a given γ > 0 and a path π with starting time σπ , the √ scaled path πn (γ ) :
[σπ /n, ∞] → [−∞, ∞] is given by πn (γ )(t) = π(nt)/( nγ ) for each n ≥ 1.
Thus, the starting time of the scaled path πn (γ ) is σπn (γ ) = σπ /n. Similarly, for the
backward path π̂ , the scaled
√ version is π̂n (γ ) : [−∞, σπ̂ /n] → [−∞, ∞] given
by π̂n (γ )(t) = π̂(nt)/( nγ ) for each n ≥ 1. For each n ≥ 1, let Xn = Xn (γ ) :=
{πn (γ ) : (x, t) ∈ V } and Xn = Xn (γ ) := {π̂n (γ ) : (x, t) ∈ V} be the collec-
(x,t) (x,t)

tions of all the nth order diffusively scaled paths and dual paths, respectively.
The closure X n (γ ) of Xn (γ ) in (
, d
) and the closure X 
n (γ ) of Xn (γ ) in
 d  ) are (H, BH ) and (H
(
, , B  ) valued random variables, respectively. Coletti,

H
Fontes and Dias [5] showed the following.

n (γ0 ) con-
T HEOREM 2.8. For γ0 := γ0 (p) as in Theorem 1.2, as n → ∞, X
verges weakly to the standard Brownian web W .

Our main result is the joint invariance principle for {(X 

n (γ0 ), X
n (γ0 )) : n ≥ 1}
considered as (H × H, BH × B  ) valued random variables.
H

T HEOREM 2.9. {(X 

n (γ0 ), X
n (γ0 )) : n ≥ 0} converges weakly to (W , W
) as
n → ∞.

We require the following propositions to prove Theorem 2.9. We say that

(x,t) (u) : u ≤ t} is a Brownian motion going back in time if W
{W (x,t) (t − s) :=
W (t + s), s ≥ 0 where {W (u) : u ≥ t} is a Brownian motion with W (t) = x.

P ROPOSITION 2.10. For any deterministic point (x, t) ∈ R2 , there exists a

sequence of paths θ̂n(x,t) ∈ Xn (γ0 ) which converges in distribution to W
(x,t) .

√
P ROOF. For any (x, t) ∈ R2 fix tn = nt and xn = max{ nγ0 x + j :
√
j ≤ 0, ( nγ0 x + j, tn ) ∈ V}. Let θ̂n(x,t) ∈ Xn (γ0 ) be the scaling of the path
π̂ (xn ,tn ) ∈ X.
1820 R. ROY, K. SAHA AND A. SARKAR

Since G is invariant under translation by lattice points and G is uniquely

determined by G , the conditional distribution of {(xn , tn ) + ĥj (0, 0) : j ≥ 0}
given√(0, 0) ∈ V is the same as that of {ĥj (xn , tn ) : j ≥ 0}. We observe that
(xn /( nγ0 ), tn /n) → (x, t) as n → ∞ almost surely. Hence, it suffices to prove
that the scaled dual path starting from (0, 0) given (0, 0) ∈ V converges in distri-
(0,0) .
bution to W
From Proposition 2.3, we see that X (0,0) = ĥj (0, 0)(1) is an L2 martingale with
j
respect to the filtration σ ({B(z,s) , U(z,s) : z ∈ Z, s ≥ −k}). Let
(0,0)  (0,0) (0,0)    
ηn (u) := sn−1 X

j

+ X 
j +1 − Xj usn2 − sj2 / sj2+1 − sj2
(0,0) − X 
(0,0) )2 ).
for u ∈ [0, ∞) and sj2 ≤ usn2 < sj2+1 , where sn2 = nj=1 E((X j j −1
We know ηn converges in distribution to a standard Brownian motion (see
Theorem 3, [4]). Since sn2 /(nγ02 ) → 1, it can be seen that supu∈[0,M] |ηn (u) −
θ̂n(0,0) (−u)| → 0 in probability for any M > 0. So by Slutsky’s theorem, we con-
clude that θ̂n(0,0) converges in distribution to a standard Brownian motion going
backward in time. 

The next result helps in estimating the probability that a direct path and a dual
path stay close to each other for some time period. Given m ∈ N and ε, δ > 0, we
define the event
Bnε = Bnε (δ, m)

:= there exist π1n , π2n , π3n ∈ Xn such that σπ1n , σπ2n ≤ 0,
 
σπ3n ≤ nδ/n, π1n (0) ∈ [−m, m], π1n (0) − π2n (0) < ε, with
        
π1n nδ/n = π2n nδ/n and π1n nδ/n − π3n nδ/n  < ε, with
   
π1n 2nδ/n = π3n 2nδ/n .

L EMMA 2.11. For any m ∈ N and ε, δ > 0, we have

 
P Bnε (δ, m) ≤ C1 (δ, m)ε,
where C1 (δ, m) is a positive constant, depending only on δ and m.

P ROOF. Let Dnε be the unscaled version of the event Bnε , that is,
  
Dnε := there exist (x, 0), (y, 0), z, nδ ∈ V such that
√ √ √
x ∈ [−m nγ0 , m nγ0 ], |x − y| < nεγ0 and hnδ (x, 0) = hnδ (y, 0),
 nδ  √  
h (x, 0)(1) − z < nεγ0 , h2nδ (x, 0) = hnδ z, nδ .
√ √
On the event Dnε there exists l ∈ [−m nγ0 , m nγ0 ] ∩ Z such that the unscaled
paths starting from (l, 0) and (l + 1, 0) (as in Figure√3) do not meet in time nδ—
an event which occurs with probability at most C2 / nδ for some constant C2 > 0
 HACK’S LAW IN A DRAINAGE NETWORK MODEL 1821

F IG . 3. The vertices (l, 0) and (l + 1, 0) and the corresponding vertex (k, nδ) as required in the
proof of Lemma 2.11.

nδ
√ Supposing h (l, 0)(1) = k, two
(see Theorem 4 of Coletti, Fontes and Dias [5]).
unscaled paths, one starting from√a vertex  nεγ0  distance to the left of k and
the other starting from a vertex  nεγ0√  distance
√ to the right of k, do not meet in
time nδ has a probability at most C2 2 nεγ0 / nδ for all k ∈ Z. Thus, summing
over all possibilities of l and k and using Markov property we have
 2m√nγ0
 


P Dnε ≤ P hnδ (l, 0)(1) = k = hnδ (l + 1, 0)(1) and
√
l=−2m nγ0 k∈Z

nδ  √   √ 
h k −  nεγ0 , nδ = hnδ k +  nεγ0 , nδ
√ √
2m nγ0
2C2 nεγ0   nδ
≤ √ P h (l, 0)(1) = k = hnδ (l + 1, 0)(1)
√ nδ
l=−2m nγ0 k∈Z
√ √
2m nγ0
2C2 nεγ0  nδ
≤ √ P h (l, 0)(1) = hnδ (l + 1, 0)(1)
√ nδ
l=−2m nγ0
√ √
2m nγ0
2C2 nεγ0 C2
≤ √ √
√ nδ nδ
l=−2m nγ0

≤ C1 (δ, m)ε. 
1822 R. ROY, K. SAHA AND A. SARKAR

P ROOF OF T HEOREM 2.9. Since X consists of noncrossing paths only, Propo-


sition 2.10 implies the tightness of the family {Xn : n ≥ 1} (see Proposition B.2 in
the Appendix of Fontes et al. [9]). The joint family {(X 
n , X
n ) : n ≥ 1} is tight
since each of the two marginal families is tight. To prove Theorem 2.9, it suffices
to show that for any subsequential limit (W , Z ) of {(X 
n , X
n ) : n ≥ 1}, the random
variable Z satisfies the conditions given in Proposition 2.7.
Consider a convergent subsequence of {(X 
n , X
n ) : n ≥ 1} such that (W , Z ) is
its weak limit and by Skorohod’s representation theorem, we may assume that the
convergence happens almost surely. For ease of notation, we denote the convergent
subsequence by itself.
From Proposition 2.10, it follows that for any deterministic (x, t) ∈ R2 there
exists a path π̂ ∈ Z starting at (x, t) going backward in time almost surely.
Since (X 
n , X
n ) converges to (W , Z ) almost surely, if a dual path in Z crosses
a path in W , there exists a dual path in Xn which crosses a path in Xn , for some
n ≥ 1, yielding a contradiction. Hence, the paths in Z do not cross paths in W
almost surely (for details, see Roy, Saha and Sarkar [27]).
Now, to prove that condition (3) in Proposition 2.7 is satisfied, we define the
following event: for δ > 0 and positive integer m ≥ 1, let

A(δ, m) := there exist paths π ∈ W and π̂ ∈ Z with σπ , σπ̂ ∈ (−m, m),
and there exists t0 such that σπ < t0 < t0 + δ < σπ̂ ,

and −m < π(t) = π̂(t) < m for all t ∈ [t0 , t0 + δ] .
It is enough to show that for any fixed δ > 0 and for m ≥ 1, we have P(A(δ,
m)) = 0.
We present here the idea of the proof; more details are available in Roy, Saha
and Sarkar [27]. Fix ε > 0. Since we are in a setup where the scaled paths converge
almost surely, for all large n there exist π1n ∈ Xn and π̂ n ∈ Xn within ε distance of
π and π̂ , respectively. Using the fact that a dual vertex lies in the middle of two
open vertices and the forward paths cannot cross the dual paths, it follows that for
all large n there exist π2n , π3n ∈ Xn such that:
(a) max{|π1n (σπ2n ) − π2n (σπ2n )|, |π1n (σπ3n ) − π3n (σπ3n )|} < 4ε;
(b) π1n (σπ2n + δ/3) = π2n (σπ2n + δ/3) and π1n (σπ3n + δ/3) = π3n (σπ3n + δ/3).
6m/δ
This gives us that A(δ, m) ⊆ lim infn→∞ j =1 Bn4ε (δ/3, 2m; j ).
Here, Bn4ε (δ/3, 2m; j ) is a translation of the event Bn4ε (δ/3, 2m), considered
in Lemma 2.11; translated such that the starting time of the paths πn1 and πn2 are
shifted by −m + j nδ/3/n (see Figure 4).
By translation invariance of our model and Lemma 2.11, for all n ≥ 1 we have
P(Bn4ε (δ/3, 2m; j )) ≤ 4C1 (δ/3, 2m)ε. This completes the proof. 
 HACK’S LAW IN A DRAINAGE NETWORK MODEL 1823

 and the approximating dashed paths

F IG . 4. The event A(δ, m). The bold paths are from (W, W)
are from (Xn , Xn ).

3. Proof of Theorem 1.2. Let ξ := ξW (0, 1) and ξn := ξXn (0, 1) be as defined

in (14). The proof of Theorem 1.2 follows from the following proposition.

P ROPOSITION 3.1. E[ξn ] → E[ξ ] as n → ∞.

We first complete the proof of Theorem 1.2 assuming Proposition 3.1.

P ROOF OF T HEOREM 1.2. Using the translation invariance of our model, we

have
√ √
 nγ0 
√   nγ0
nγ0 P L(0, 0) > n = E(1{L(k,n)>n} ) × √
k=0
 nγ0  + 1
√
nγ0 1
= E(ξn ) × √ → E(ξ ) = √ as n → ∞.
 nγ0  + 1 π
This proves Theorem 1.2. 

Proposition 3.1 will be proved through a sequence of lemmas.

To state the next lemma, we recall from Theorem 2.9 that (X 
n , X
n ) ⇒ (W , W
)
as n → ∞. Using Skorohod’s representation theorem, we assume that we are
working on a probability space where dH×H ((X 
n , X
n ), (W , W
)) → 0 almost
surely as n → ∞.
1824 R. ROY, K. SAHA AND A. SARKAR

L EMMA 3.2. For t1 > t0 , we have

 
P ξXn (t0 , t1 ) = ξW (t0 , t1 ) for infinitely many n = 0.

P ROOF. We prove the lemma for t0 = 0 and t1 = 1, that is, for ξn = ξXn (0, 1)
and ξW (0, 1), the proof for general t0 , t1 being similar. First, we show that, for all
k ≥ 0,
(17) lim inf 1{ξn ≥k} ≥ 1{ξ ≥k} almost surely.
n→∞

Indeed, for k = 0, both 1{ξn ≥k} and 1{ξ ≥k} equal 1. For k ≥ 1, (17) follows from
almost sure convergence of (X 
n , X
n ) to (W , W
) and from the properties of the set
MW (0, 1) as described in Proposition 2.5.
To complete the proof, we need to show that P(lim supn→∞ {ξn > ξ }) = 0. This
is equivalent to showing that P(k0 ) = 0 for all k ≥ 0, where

k0 := ω : ξn (ω) > ξ(ω) = k for infinitely many n .
Consider k = 0 first. From Proposition 2.5, it follows that on the event ξ = 0,
almost surely we can obtain γ := γ (ω) > 0 such that MW (0, 1) ∩ (−γ , 1 + γ ) =
∅. From the almost sure convergence of (X 
n , X
n ) to (W , W
), we have P(0 ) = 0.
0
For k > 0, on the event 0 we show a forward path π ∈ W coincides with a
k

dual path π̂ ∈ W  for a positive time which leads to a contradiction. From Propo-
sition 2.5, it follows that given η > 0, there exist m0 ∈ N and s0 ∈ (1/m0 , 1) such
that P(ξW (1/m0 , 1) = ξW (1/m0 , s0 ) = ξW (0, 1) = k) > 1 − η, that is, the paths
leading to any single point considered in MW (0, 1) = MW (1/m0 , 1) have coa-
lesced before time s0 . Fix 0 < ε < 1/m0 such that (x − ε, x + ε) ⊂ (0, 1) for all
x ∈ MW (1/m0 , 1) and the ε-tubes around the k paths contributing to MW (s0 , 1),
viz., π1 (t), . . . , πk (t), t ∈ [s0 , 1], given by

Tεi := (x, t) : πi (t) − ε ≤ x ≤ πi (t) + ε, s0 ≤ t ≤ 1 for i = 1, . . . , k,
are disjoint. Since we have almost sure convergence on the event k0 , there exists
n n
n0 such that one of the k tubes must contain at least two paths, π1 0 , π2 0 (say)
of Xn0 which do not coalesce by time 1. From the construction of dual paths, it
 1+
 lying between π n0 and
follows that there exists at least one dual path π̂ n0 ∈ X n0 1
n
π2 0 for t ∈ [s0 , 1], and hence we must have an approximating π̂ ∈ W 1+ close
to π̂ n0 for t ∈ [s0 , 1]. Since we have only finitely many disjoint k tubes, taking
ε → 0 and using compactness of W  we obtain that there exists π̂ ∈ W
 such that
π̂(t) = πi (t) for t ∈ [s0 , 1] and for some 1 ≤ i ≤ k. This violates the property of
Brownian web and its dual that they do not spend positive Lebesgue time together.
Hence, P(k0 ) = 0 for all k ≥ 0 and this completes the proof of the lemma. 

Lemma 3.2 immediately gives the following corollary.

 HACK’S LAW IN A DRAINAGE NETWORK MODEL 1825

C OROLLARY 3.2.1. As n → ∞, ξn converges in distribution to ξ .

Corollary 3.2.1 along with the following lemma completes the proof of Propo-
sition 3.1.

L EMMA 3.3. The family {ξn : n ∈ N} is uniformly integrable.

P ROOF. For m ∈ N, let

 Km
Km = [−m, m]2 ∩ Z2 and m := (0, 1), (0, −1), (1, 1), (1, −1) .
We assign the product probability measure P whose marginals for u ∈ Km are
given by
⎧
⎪ p,
⎪ for a = 1 and b ∈ {1, −1},
 ⎨
2
P ζ : ζ (u) = (a, b) =
⎪
⎪ (1 − p)
⎩ , for a = 0 and b ∈ {1, −1}.
2
P is the measure induced by the random variables {(Bu , Uu ) : u ∈ Km }.
For ζ ∈ m and for K ⊆ Km , the K cylinder of ζ is given by C(ζ, K) := {ζ  :

ζ (u) = ζ (u) for all u ∈ K}. For any two events A, B ⊆ m , let

AB := ζ : there exists K = K(ζ ) ⊆ Km such that C(ζ, K) ⊆ A,
 
and C ζ, K  ⊆ B for K  = Km \ K
denote the disjoint occurrence of A and B. Note that this definition is associative,
that is, for any A, B, C ⊆ m we have (AB)C = A(BC).
Let
 √
Fnm := there exist (u1 , n), (u2 , n) ∈ V with 0 ≤ u1 < u2 ≤ nγ0 and
 l   l 
v1 , l , v2 , l ∈ V for all 0 ≤ l ≤ n such that

−m ≤ v1l < ĥl (u1 , n)(1) < ĥl (u2 , n)(1) < v2l ≤ m ,

Enm (k) := for 1 ≤ i = j ≤ k, there exists (xi , 0) ∈ V with
√
hn (xi , 0)(1) ∈ [0, nγ0 ], hn (xi , 0) = hn (xj , 0), hl (xi , 0)(1) ∈ [−m, m]

for all 0 ≤ l ≤ n .
We claim that for all k ≥ 2,
(18) Enm (3k) ⊆ Fnm Fnm  · · · Fnm .
! "# $
k times

We prove it for k = 2. For general k, the proof is similar. Let (ui , n) ∈ V, 1 ≤
i ≤ 5 and (xi , 0) ∈ V , 1 ≤ i ≤ 6 be as in Figure 5. The region explored to obtain the
1826 R. ROY, K. SAHA AND A. SARKAR

F IG . 5. The event Enm (6).

vertex ĥj (ui , n) for 1 ≤ j ≤ n is contained in n−1

l=0 [h (xi , 0)(1), h (xi+1 , 0)(1)] ×
l l

{l}. Thus, the regions explored to obtain the dual paths starting from (u1 , n), (u2 , n)
and the dual paths starting from (u4 , n), (u5 , n) are disjoint (see Figure 5). Hence,
it follows that Enm (6) ⊆ Fnm Fnm .
Since the event Enm (k) is monotonic in m, from (18) we get
% &  
P(ξn ≥ 3k) = P lim Enm (3k) = lim P Enm (3k)
m→∞ m→∞
   
≤ lim P Fnm  · · · Fnm = lim P Fnm  · · · Fnm .
m→∞ m→∞

Applying the BKR inequality (see Reimer [24]), we get

  k % % &&k  k
(19) P(ξn ≥ 3k) ≤ lim P Fnm = P lim Fnm = P(Fn ) ,
m→∞ m→∞
√
where Fn := {there exist (u1 , n), (u2 , n) ∈ V with 0 ≤ u1 < u2 ≤ nγ0 such that
ĥn (u1 , n) = ĥn (u2 , n)}. √
For any (x, t) ∈ R2 fix tn = nt and xn = max{ nγ0 x + j : j ≤ 0,
√
( nγ0 x + j, tn ) ∈ V}. Let θ̂n(x,t) ∈ Xn (γ0 ) be the scaling of the path π̂ (xn ,tn ) ∈ X.
Define

Fn := θ̂n(0,1) and θ̂n(1,1) do not coalesce in time 1 .
We observe that Fn ⊆ Fn . Now P(Fn ) converges to the probability that two
independent Brownian motions starting at a distance 1 from each other do not
meet by time 1. Since limn→∞ P(Fn ) < 1, the family {ξn : n ∈ N} is uniformly
integrable. 
 HACK’S LAW IN A DRAINAGE NETWORK MODEL 1827

R EMARK 3.4. It is to be noted that Newman, Ravishankar and Sun [23] also
used ideas of negative correlation to establish the weak convergence of MXn as a
point process on R for a more general setup where paths can cross each other. In
our case, the negative correlation ideas come in a much less essential manner only
to establish uniform integrability as the noncrossing nature of paths enable us to
obtain Corollary 3.2.1.

4. Proofs of Theorems 1.3 and 1.4. In this section, we prove Theorems 1.3
and 1.4. The main idea of the proof is that the horizontal distance between the
ˆ
dual paths π̂ r̂(x,t) and π̂ l(x,t) (see Figure 6) form a Brownian excursion process
after scaling. The cluster C(x, t) being enclosed between these two paths, its size
is related to the area under the Brownian excursion.
+
For a formal proof, we need to introduce some notation. For τ > 0, let S τ , S τ :
C[0, ∞) → R be defined by S (f ) := inf{t ≥ 0 : f (t +s) ≥ f (t) for all 0 ≤ s ≤ τ }
τ
+ +
and S τ (f ) := inf{t ≥ 0 : f (t +s) > f (t) for all 0 < s ≤ τ }. Let T τ : C[0, ∞) →
C[0, ∞) be the map given by
  +   + +
T τ (f )(s) := f S + s − f S if S τ < ∞,
+ τ τ ,
(20)
f (s), otherwise.
+
For a Brownian motion W with W (0) = 0, we define W τ = T τ (W ). From
+
Bolthausen [3], we have S τ = S τ < ∞ almost surely under the measure induced
by W on C[0, ∞) and W 1 |[0,1] = W + where W + is the standard Brownian me-
d

ander process defined in (3). From the scaling property of Brownian motion, it

ˆ
F IG . 6. The two dual paths π̂ l(x,t) and π̂ r̂(x,t) enclose the cluster C(x, t). These dual paths after
scaling are each Brownian paths.
1828 R. ROY, K. SAHA AND A. SARKAR

d √
follows that {W τ (s) : s ∈ [0, τ ]} = { τ W + (s/τ ) : s ∈ [0, τ ]}. Durrett, Iglehart
and Miller [7] (Theorem 2.1) proved that W |1{mins∈[0,1] W (s)≥−ε} ⇒ W + as ε ↓ 0.
Using this result and the scaling property of W τ , given above, straightforward cal-
culations imply the following lemma and its corollary (for details, see Roy, Saha
and Sarkar [27]).

L EMMA 4.1. For τ > 0 considering W as a standard Brownian motion on

[0, ∞) starting from 0, we have W |1{mint∈[0,τ ] W (t)≥−1/n} ⇒ W τ as n → ∞.

'τ as the process on C[0, ∞) given by

Define W

'τ (t) := W τ (t), if 0 ≤ t ≤ τ ,
W '(t − τ ),
W τ (τ ) + W otherwise,
' is a Brownian motion on [0, ∞), independent of W τ , with W
where W '(0) = 0. For
f ∈ C[0, ∞), let tf := inf{s > 0 : f (s) = 0} with tf = ∞ if f (s) = 0 for all s > 0.
Consider the mapping H : C[0, ∞) → C[0, ∞) given by H (f )(t) := 1{t≤tf } f (t).
We define W +,τ = H (W τ ). A similar argument as that of Lemma 4.1 gives us the
following corollary.

C OROLLARY 4.1.1. 'τ and W +,τ = H (W

d
For τ > 0, we have, W τ = W 'τ ).
d

Let A ⊂ C[0, ∞) be such that


A := f ∈ C[0, ∞) : tf < ∞ and for every ε > 0 there exists
(21)
s ∈ (tf , tf + ε) with f (s) < 0 .
From Corollary 4.1.1, it follows that P(W τ ∈ A) = 1. Hence, H is continuous
almost surely under the measure induced W τ on C[0, ∞).
 ∞ by+,τ
Next, we obtain the distribution of 0 W (t) dt.

L EMMA 4.2. For τ, λ > 0, we have

 ∞  √  ∞
τ  
P W +,τ
(t) dt > λ =  + λt −3/2 dt.
t −3/2 F I0
0 2 τ
P ROOF. We give here a straightforward proof using random walk. Let {Sn :
n ≥ 0} be a symmetric random walk with variance 1 starting at S0 = 0. Since
P(W τ ∈ A) = 1, minor modification of the argument used to prove Lemmas 2.4
+
and 2.5, Bolthausen [3] shows that H ◦ T τ is almost surely continuous under the
measure induced by W on C[0, ∞) (for details, see Roy, Saha and Sarkar [27]).
From Donsker’s invariance principle and from  the continuous mapping theorem, it
follows that for λ > 0, a continuity point of 0∞ W +,τ (t) dt, we have
 ∞   ∞ 
 + 
P W +,τ (t) dt > λ = lim P H T τ (Yn ) (t) dt > λ ,
0 n→∞ 0
 HACK’S LAW IN A DRAINAGE NETWORK MODEL 1829

where
Sk (nt − [nt]) k k+1
(22) Yn (t) := √ + √ (Sk+1 − Sk ) for ≤t < .
n n n n
A similar argument as in Lemma 3.1 of Bolthausen [3] gives us that (for details,
see Roy, Saha and Sarkar [27])
 ∞ 
 τ+ 
P H T (Yn ) (t) dt > λ
0
 ∞  

=P H (Yn )(t) dt > λ min Yn (t) ≥ 0, t0 > nτ ,
0 t∈[0,τ ]

where t0 := inf{n > 0 : Sn = 0} is the first return time to 0 of the random walk.
Hence for λ > 0, a continuity point of W +,τ , we obtain
 ∞ 
+,τ
P W (t) dt > λ
0
 ∞  

= lim P H (Yn )(t) dt > λ min Yn (t) ≥ 0, t0 > nτ
n→∞ 0 t∈[0,τ ]
∞
 n3/2 P(t0 = nτ  + j )
= lim √
n→∞ n( nP(t0 > nτ ))
j =1
 ∞  

×P H (Yn )(t) dt > λ min Yn (t) ≥ 0, t0 = nτ  + j
0 t∈[0,τ ]
 ∞
1
= lim √ gn (t)fn (t) dt,
n→∞ nP(t0 > nτ ) nτ /n

where for t ≥ nτ /n, fn (t) = P( 0∞ H (Yn )(u) du > λ| mint∈[0,τ ] Yn (t) ≥ 0, t0 =
nt + 1) and gn (t) = n3/2 P(t0 = nt + 1). It is known that (see Kaigh [16])
(
√ 2 1
lim nP(t0 > n) = and lim n3/2 P(t0 = n) = √ .
n→∞ π n→∞ 2π
Hence, from Theorem 2.6 Kaigh [16] together with the continuous mapping theo-
 ∞ +,τ
rem and
√ 
the scaling property of the Brownian motion we have P( 0 W (t) dt >
τ ∞ −3/2  −3/2 +
λ) = 2 τ t FI + (λt ) dt. Finally, I0 being a continuous random variable
0 
(see Louchard and Janson [20]), it follows that the random variable 0∞ W +,τ (t) dt
is continuous. This completes the proof. 

ˆ t) denote the right and

4.1. Proof of Theorem 1.3. Recall that r̂(x, t) and l(x,
left dual neighbours, respectively, of (x, t) ∈ V . Let D̂k (x, t) := ĥk (r̂(x, t))(1) −
ˆ t))(1) where ĥ is as defined after (11). Consider the continuous function
ĥk (l(x,
1830 R. ROY, K. SAHA AND A. SARKAR

(x,t)
D̂n ∈ C[0, ∞) given by
D̂k (x, t) (ns − [ns])  
D̂n(x,t) (s) := √ + √ D̂k+1 (x, t) − D̂k (x, t)
γ0 n γ0 n
(23)
k k+1
for ≤s≤ .
n n
Fix τ > 0. For an H × H  valued random variable (K, K)
 and for x ∈ MK (0, τ )
(x,τ )
let π̂r be defined as

τ + with x < π̂1 (τ ) < π̂(τ ),
if σπ̂ = τ and there is no π̂1 ∈ K
π̂r(x,τ ) := π̂,
π̂0 , otherwise,
(x,τ )
where π̂0 denotes the constant zero function with σπ̂0 = τ . In other words, π̂r ∈
 +
K is such that among all π̂ ∈ K , π̂r
τ  τ + (x,τ )
(τ ) is closest to (x, τ ) on the right.
(x,τ )
Similarly, π̂l is defined as the path closest to (x, τ ) on the left.
 with σπ̂ ≥ τ , let g(π̂ ) ∈ C[0, ∞) be given by g(π̂ )(t) := π̂(τ − t)
For π̂ ∈

for t ≥ 0. Fix f ∈ Cb [0, ∞) and define

     (x,τ ) 
 (τ, f ) :=
κ(K,K) f g π̂r(x,τ ) − g π̂l .
x∈MK (0,τ )

) (τ, f ), and κn (τ, f ) := κ(X

Let κ(τ, f ) := κ(W ,W n ) (τ, f ). Comparing with the

n ,X
definitions introduced in (14), for mf = sup{|f (s)| : s ∈ [0, ∞)} we have
(24) κ(τ, f ) ≤ mf ξW (0, τ ), κn (τ, f ) ≤ mf ξXn (0, τ ) for all n ≥ 1.
(x,τ )
From Proposition 2.5, we know that for each x ∈ MW (0, τ ), there exist π̂r ,
(x,τ )
π̂l  (x,τ )
∈ W both starting from (x, τ ) with π̂r (x,τ )
(0) > π̂l (0).
The following lemma is the main tool for establishing Theorem 1.3 and Theo-
rem 1.4.

L EMMA 4.3. For τ > 0 and f ∈ Cb [0, ∞), we have

 
(25) lim E κn (τ, f ) = E κ(τ, f ) .
n→∞

P ROOF. From (24) and Lemma 3.3, it follows that the family {κn (τ, f ) : n ∈
N} is uniformly integrable. Hence, it suffices to show that κn (τ, f ) converges in
distribution to κ(τ, f ) as n → ∞. We assume that we are working on a probability
space such that (X 
n , X
n ) converges to (W , W
) almost surely in (H × H , d  ).
H ×H
From Lemma 3.2, we have limn→∞ ξXn (0, τ ) = ξW (0, τ ) almost surely, and hence
from (24) for ξW (0, τ ) = 0, we have κn (τ, f ) = κ(τ, f ) = 0 for all n large.
Next, we consider the case ξW (0, τ ) = k ≥ 1. Suppose MW (0, τ ) = {x1 , . . . , xk }.
From Lemma 3.2, we have that MXn (0, τ ) = {x1n , . . . , xkn } for all large n and
limn→∞ xin = xi for all 1 ≤ i ≤ k. Fix T ≥ 0. To complete the proof, it is enough to
 HACK’S LAW IN A DRAINAGE NETWORK MODEL 1831

(x ,τ ) (x n ,τ ) (x ,τ ) (x n ,τ )
show that sup{|π̂r i (τ − s) − π̂r i (τ − s)| ∨ |π̂l i (τ − s) − π̂l i (τ − s)| :
s ∈ [0, τ + T ]} → 0 as n → ∞ for all 1 ≤ i ≤ k.
We observe that for yi ∈ (π̂r(xi ,τ ) (0), π̂l(xi ,τ ) (0)) ∩ Q there exists π (yi ,0) ∈ W
such that π (yi ,0) (τ ) = xi . We choose ε = ε(ω) > 0 so that for all 1 ≤ i ≤ k:
(a) (xi − ε, xi + ε) ⊂ (0, 1), (xi − 2ε, xi + 2ε) ∩ MW (0, τ ) = {xi } and
(b) (π̂r(xi ,τ ) (0) − π (yi ,0) (0)) ∧ (π (yi ,0) (0) − π̂l(xi ,τ ) (0)) > 2ε.
Let n0 = n0 (ω) be such that, for all n ≥ n0 :
(i) ξXn (0, τ ) = ξW (0, τ ) and
 τ+
(ii) for all 1 ≤ i ≤ k there exist π̂i1,n , π̂i2,n ∈ X 
n and πi ∈ Xn
n 0− such that
1,n (xi ,τ ) 2,n (xi ,τ )
sup{|π̂i (τ − s) − π̂r (τ − s)| ∨ |π̂i (τ − s) − π̂l (τ − s)| ∨ |πin (τ − s) −
π (yi ,0) (τ − s)| : s ∈ [0, τ + T ]} < ε.
The choice of n0 ensures that MXn (0, τ ) ∩ (xi − ε, xi + ε) = {xin }. Since there
exist only two dual paths starting from (xi , τ ), because of the uniqueness of xin in
the interval (xi − ε, xi + ε) and the noncrossing nature of our paths we must have
(x n ,τ ) (x n ,τ )
π̂r i (τ − s) = π̂i1,n (τ − s) and π̂l i (τ − s) = π̂i2,n (τ − s) for all s ∈ [0, τ + T ]
and for all n ≥ n0 (for details, see Roy, Saha and Sarkar [27]). Since T ≥ 0 is
chosen arbitrarily, this completes the proof. 

The next lemma calculates E[κ(τ, f )].

L EMMA 4.4. For τ > 0 and f ∈ Cb [0, ∞), we have

  √  √
E κ(τ, f ) = E f 2W +,τ / πτ .

P ROOF. Let In ⊂ {0, 1, . . . , n − 1} given by In := {i : 0 ≤ i ≤ n − 1, π̂ (i/n,τ ) ,

 such that π̂ (i/n,τ ) (0) < π̂ ((i+1)/n,τ ) (0)}. We define
π̂ ((i+1)/n,τ ) ∈ W
   
Rn (τ, f ) = f g π̂ ((i+1)/n,τ ) − π̂ (i/n,τ ) .
i∈In

From Proposition 2.5, we know MW (0, τ ) ∩ Q = ∅. For each x ∈ MW (0, τ ),

set lnx = nx/n and rnx = lnx + (1/n). Since there are exactly two dual paths
π̂r(x,τ ) and π̂l(x,τ ) starting from (x, τ ) with π̂r(x,τ ) (0) > π̂l(x,τ ) (0), from Proposi-
x x
tion 3.2(e) of Sun and Swart [29] it follows that {π̂ (ln ,τ ) : n ∈ N} and {π̂ (rn ,τ ) :
n ∈ N} converge to π̂l
(x,τ )
and π̂r
(x,τ )  d  ) as n → ∞.
, respectively, in (
,

Hence, Rn (τ, f ) → κ(τ, f ) almost surely as n → ∞. For each i ∈ In , there ex-

ist yi ∈ (π̂ (i/n,τ ) (0), π̂ ((i+1)/n,τ ) (0)) ∩ Q and π (yi ,0) ∈ W such that π (yi ,0) (τ ) ∈
MW (0, τ ). Hence, for mf = sup{|f (t)| : t ≥ 0} we have Rn (τ, f ) ≤ mf ξW (0, τ )
for all n. As E[ξW (0, τ )] < ∞, the family {Rn (τ, f ) : n ∈ N} is uniformly inte-
grable, and hence we have limn→∞ E[Rn (τ, f )] = E[κ(τ, f )]. From the fact that
1832 R. ROY, K. SAHA AND A. SARKAR

d √
g(π̂ ((i+1)/n,τ ) ) − g(π̂ (i/n,τ ) ) = H (1/n + 2W ) where W denotes the standard
Brownian motion on [0, ∞), we have

lim E Rn (τ, f )
n→∞
)  √  √ *
= lim nE f H (1/n + 2W ) 1/n + min 2W (t) > 0
n→∞ t∈[0,τ ]
% √ &
× P 1/n + min 2W (t) > 0
t∈[0,τ ]
)  √  √ *
= lim E f H (1/n + 2W )  min 2W (t) > −1/n n
n→∞ t∈[0,τ ]
 √ 
× 2(1/ 2τ n) − 1
 √  √
= E f 2W +,τ / πτ ,
where the last equality follows from Lemma 4.1, Slutsky’s theorem and continuous
mapping theorem. This completes the proof. 

Now, to complete the proof of Theorem 1.3 we need the following lemmas.
(0,0) √ +,τ
L EMMA 4.5. For τ > 0, we have D̂n |1{L(0,0)>nτ } ⇒ 2W as n → ∞.

P ROOF. Using translation invariance of our model, we have

    E[κn (τ, f )] E[κ(τ, f )]  √ 
E f D̂n(0,0) |1{L(0,0)>nτ } = → = E f 2W +,τ .
E[ξXn (0, τ )] E[ξW (0, τ )]
This holds for all f ∈ Cb [0, ∞) which completes the proof. 

L EMMA 4.6. For τ > 0, we have:

(0,0) (0,0) P
(a) sup{|D̂n (s) − Dn (s)| : s ≥ 0}|1{L(0,0)>nτ } → 0 as n → ∞,
P
(b) sup{|Kn(0,0) (s) − pDn(0,0) (s)| : s ≥ 0}|1{L(0,0)>nτ } → 0 as n → ∞.

P ROOF. For part (a), fix 0 < α < 1/2, T ≥ 0 and we observe that
   
P sup D̂k (0, 0) − Dk (0, 0) : k ≥ 0 ≥ nα , L(0, 0) > nτ
  
≤ P max D̂k (0, 0) − Dk (0, 0) : 0 ≤ k ≤ n(τ + T ) + 1 ≥ nα ,
  
L(0, 0) > nτ + P L(0, 0) > n(τ + T ) .
√
Because of Theorem 1.2, it is enough to show that nP(max{|D̂k (0, 0) −
Dk (0, 0)| : 0 ≤ k ≤ n(τ + T ) + 1} ≥ nα , L(0, 0) > nτ ) → 0 as n → ∞. Here,
we present the simple idea behind the proof; the details are available in Roy, Saha
and Sarkar [27].
 HACK’S LAW IN A DRAINAGE NETWORK MODEL 1833

The distance dkl between lk (0, 0) and the closest open vertex to the left of
α
lk (0, 0) being nα or more has a probability (1 − p)n . Thus, the probability that
the maximum such difference for 0 ≤ k ≤ n(τ + T ) + 1 is bigger that nα is
α
of the order n(1 − p)n . Similarly, for the distance dkr associated with the ver-
tex r (0, 0). Since |D̂k (0, 0) − Dk (0, 0)| ≤ dkl + dkr , as n → ∞, we have that
√ k
nP(max{|D̂k (0, 0) − Dk (0, 0)| : 0 ≤ k ≤ n(τ + T ) + 1} ≥ nα , L(0, 0) > nτ ) con-
verges to 0. √
For part (b) of the lemma, we need Dn(0,0) |1{L(0,0)>nτ } ⇒ 2W +,τ as n → ∞
which√ follows from part (a) and Lemma 4.5. Hence, rk (0, 0) − lk (0, 0) is of the or-
der n. Also given lk (0, 0) and rk (0, 0), the number of open vertices lying between
these vertices has a binomial distribution with parameters (rk (0, 0) − lk (0, 0) − 1)
and p. Since these open vertices together with lk (0, 0) and rk (0, 0) constitute
Ck (0, 0), the proof follows from similar order comparisons as done in (a). 

P ROOF OF T HEOREM 1.3. We remarked that W 1 |[0,1] = W +,1 |[0,1] = W + .

d

The proof of Theorem 1.3 follows from Lemmas 4.5 and 4.6 and Slutsky’s theorem
with the choice of τ = 1. 
√
4.2. Proof of Theorem 1.4. For λ > 0, let λ̄ := λ3/2 ( 2γ0 p)−1 . We show that:

L EMMA 4.7. For τ, λ > 0,

 ∞

√ 
lim nP L(0, 0) > nτ, #Ck (0, 0) > (λn)3/2
n→∞
k=0
√  ∞ 
1 +,τ
= √ P 2 W (t) dt > λ̄
γ0 πτ 0
 ∞
1  
= √  + λ̄t −3/2 t −3/2 dt.
F I
2γ0 π τ 0


P ROOF. For f ∈ C[0, ∞) let I (f ) := 0∞ H (f )(t) dt. Since P(W τ ∈ A) = 1
where A is defined as in (21), I is almost surely continuous under the measure
induced by W τ on C[0, ∞). The proof follows from Theorem 1.3(ii) and the con-
tinuous mapping theorem. 

From the previous lemma, we derive the following.

C OROLLARY 4.7.1. For λ > 0, we have

 ∞
√   1  
lim nP #C(0, 0) > (λn)3/2 = √  + λ̄t −3/2 t −3/2 dt.
F
n→∞ I
2γ0 π 0 0
1834 R. ROY, K. SAHA AND A. SARKAR

P ROOF. For any τ > 0, we have P(#C(0, 0)√ > (nλ) ) ≥ P(L(0,
3/2 0) >
nτ, #C(0, 0) > (nλ) ), and hence lim infn→∞ nP(#C(0, 0) > (nλ)3/2 ) ≥
3/2
1√  ∞ 
2γ π 0
FI + (λ̄t −3/2 )t −3/2 dt.
0 0
We observe that
√  
nP L(0, 0) ≤ nτ, #C(0, 0) > (nλ)3/2
nτ  
√ 
≤ nP D̂k (0, 0) > (nλ) 3/2

k=0
+nτ  ,
√ 
≤ nE k (0, 0) (nλ)−3/2
D
k=0
√    
= n nτ  + 1 E D0 (0, 0) (nλ)−3/2 ,

ˆ 0))(1) : k ≥
where we have used the fact that {D̂k (0, 0) = ĥk (r̂(0, 0))(1) − ĥk (l(0,
0} is a martingale (see Proposition 2.3). From the earlier discussions, it also follows
that E(D 0 (0, 0)) ≤ 2E(G) = 2(1 − p)p −1 where G is a geometric random vari-
√
able. Thus, lim supn→∞ nP(L(0, 0) ≤ nτ, #C(0, 0) > (nλ)3/2 ) = 0 as τ → 0,
which completes the proof. 

P ROOF OF T HEOREM 1.4. We first recall the result Lemma 6.1 of Resnick
[25], page 174 which states that for nonnegative Radon measures μ, μn , n ≥ 1,
v
on [0, ∞)d \ {0} we have μn → μ if and only if μn ([0, x1 ] × · · · × [0, xd ])c →
μ([0, x1 ] × · · · × [0, xd ])c for all x1 , . . . , xd ≥ 0 with (x1 , . . . , xd ) = 0. This result
implies that Lemma 4.7 together with Corollary 4.7.1 and Theorem 1.2 prove (6).
Fix τ > 0, λ > 0. For α < 2/3, δ > 0 and for all large n, we have P(L(0, 0) >
nτ, #C(0, 0) > (nλ)1/α ) ≤ P(L(0, 0) > nτ, √ #C(0, 0) > (nδ)3/2 ). Fix any ε > 0
 ∞ +,τ
and choose δ = δ(ε) > 0 so that γ √π τ P( 2 0 W (t) dt > 
1
δ ) < ε, where
0
 −1
δ = δ (γ0 p) . From Lemma 4.7, we have
3/2

√  
lim sup nP L(0, 0) > nτ, #C(0, 0) > (nλ)1/α < ε.
n→∞

On the other hand, from the properties of W + and W τ , it follows that

∞
P( 0 W +,τ (t) dt > 0) = 1 for τ > 0. Now for α > 2/3 and δ > 0 we have
P(L(0, 0) > nτ, #C(0, 0) > (nλ)1/α ) ≥ P(L(0, 0) > nτ, #C(0, 0) > (nδ)3/2 ) for
all large n. Again from Lemma 4.7, we have
√  
lim inf nP L(0, 0) > nτ, #C(0, 0) > (nλ)1/α
n→∞
√  ∞ 
1
≥ √ P 2 +,τ 
W (t) dt > δ .
γ0 πτ 0
 HACK’S LAW IN A DRAINAGE NETWORK MODEL 1835

Since
√  
lim sup nP L(0, 0) > nτ, #C(0, 0) > (nλ)1/α
n→∞
√   1
≤ lim nP L(0, 0) > nτ = √ ,
n→∞ γ0 πτ
√
letting δ → 0, we have limn→∞ nP(L(0, 0) > nτ, #C(0, 0) > (nλ)1/α ) = γ √1π τ
0
for α > 2/3. This completes the proof of (8).
The argument for (L(0, 0), (Dmax (0, 0))1/2 ) being similar is omitted. 

Acknowledgements. Kumarjit Saha is grateful to the Indian Statistical Insti-

tute for a fellowship to pursue his Ph.D. The authors also thank the referee for
comments which led to a significant improvement of the paper.

REFERENCES
[1] A RRATIA , R. (1981). Coalescing Brownian motions and the voter model on Z. Unpublished
partial manuscript.
[2] A RRATIA , R. A. (1979). Coalescing Brownian motions on the line. ProQuest LLC, Ann Arbor,
MI. Ph.D. Thesis, Univ. of Wisconsin—Madison. MR2630231
[3] B OLTHAUSEN , E. (1976). On a functional central limit theorem for random walks conditioned
to stay positive. Ann. Probab. 4 480–485. MR0415702
[4] B ROWN , B. M. (1971). Martingale central limit theorems. Ann. Math. Stat. 42 59–66.
MR0290428
[5] C OLETTI , C. F., F ONTES , L. R. G. and D IAS , E. S. (2009). Scaling limit for a drainage
network model. J. Appl. Probab. 46 1184–1197. MR2582714
[6] D URRETT, R. T. and I GLEHART, D. L. (1977). Functionals of Brownian meander and Brown-
ian excursion. Ann. Probab. 5 130–135. MR0436354
[7] D URRETT, R. T., I GLEHART, D. L. and M ILLER , D. R. (1977). Weak convergence to Brow-
nian meander and Brownian excursion. Ann. Probab. 5 117–129. MR0436353
[8] F ERRARI , P. A., F ONTES , L. R. G. and W U , X.-Y. (2005). Two-dimensional Poisson trees
converge to the Brownian web. Ann. Inst. Henri Poincaré Probab. Stat. 41 851–858.
MR2165253
[9] F ONTES , L. R. G., I SOPI , M., N EWMAN , C. M. and R AVISHANKAR , K. (2004). The Brow-
nian web: Characterization and convergence. Ann. Probab. 32 2857–2883. MR2094432
[10] G ANGOPADHYAY, S., ROY, R. and S ARKAR , A. (2004). Random oriented trees: A model of
drainage networks. Ann. Appl. Probab. 14 1242–1266. MR2071422
[11] G IACOMETTI , A., M ARTITAN , A., R IGON , R., R INALDO , A. and RODRIGUEZ -I TURBE , I.
(1996). On Hack’s law. Water Resour. Res. 32 3367–3374.
[12] G RAY, D. M. (1961). Interrelationships of watershed characteristics. J. Geophys. Res. 66 1215–
1233.
[13] H ACK , J. T. (1957). Studies of longitudinal stream profiles in Virginia and Maryland. Report,
U.S. Geol. Surv. Prof. Pap. 294-B.
[14] H ORTON , R. T. (1945). Erosional development of streams and their drainage basins: Hydro-
physical approach to quantitative morphology. Geol. Soc. Am. Bull. 56 275–370.
[15] H URST, H. (1927). The Lake Plateau Basin of the Nile. Government Press, Cairo.
[16] K AIGH , W. D. (1976). An invariance principle for random walk conditioned by a late return to
zero. Ann. Probab. 4 115–121. MR0415706
1836 R. ROY, K. SAHA AND A. SARKAR

[17] L ANGBEIN , W. B. (1947). Topographic characteristics of drainage basins. Report, U.S. Geol.
Surv. Prof. Pap. 968-C.
[18] L EOPOLD , L. B. and L ANGBEIN , W. B. (1962). The concept of entropy in landscape evolu-
tion. Report, U.S. Geol. Surv. Prof. Pap. 500-A.
[19] L EVIN , D. A., P ERES , Y. and W ILMER , E. L. (2009). Markov Chains and Mixing Times.
Amer. Math. Soc., Providence, RI. MR2466937
[20] L OUCHARD , G. and JANSON , S. (2007). Tail estimates for the Brownian excursion area and
other Brownian areas. Electron. J. Probab. 12 1600–1632. MR2365879
[21] M ANDELBROT, B. B. (1982). The Fractal Geometry of Nature. W. H. Freeman, New York.
MR0665254
[22] M ULLER , J. E. (1973). Re-evaluation of the relationship of master streams and drainage basins:
Reply. Geol. Soc. Am. Bull. 84 3127–3130.
[23] N EWMAN , C. M., R AVISHANKAR , K. and S UN , R. (2005). Convergence of coalescing non-
simple random walks to the Brownian web. Electron. J. Probab. 10 21–60. MR2120239
[24] R EIMER , D. (2000). Proof of the van den Berg–Kesten conjecture. Combin. Probab. Comput.
9 27–32. MR1751301
[25] R ESNICK , S. I. (2007). Heavy-Tail Phenomena: Probabilistic and Statistical Modelling.
Springer, New York.
[26] RODRIGUEZ -I TURBE , I. and R INALDO , A. (1997). Fractal River Basins: Chance and Self-
Organization. Cambridge Univ. Press, New York.
[27] ROY, R., S AHA , K. and S ARKAR , A. (2015). Hack’s law in a drainage network model:
A Brownian web approach. Preprint. Available at arXiv:1501.01382.
[28] S OUCALIUC , F., T ÓTH , B. and W ERNER , W. (2000). Reflection and coalescence between
independent one-dimensional Brownian paths. Ann. Inst. Henri Poincaré Probab. Stat. 36
509–545. MR1785393
[29] S UN , R. and S WART, J. M. (2008). The Brownian net. Ann. Probab. 36 1153–1208.
MR2408586
[30] T ÓTH , B. and W ERNER , W. (1998). The true self-repelling motion. Probab. Theory Related
Fields 111 375–452. MR1640799

T HEORETICAL S TATISTICS AND M ATHEMATICS U NIT

I NDIAN S TATISTICAL I NSTITUTE
7 S. J. S. S ANSANWAL M ARG
N EW D ELHI 110016
I NDIA
E- MAIL : [email protected]
[email protected]
[email protected]