Journal of Mathematical Economics 45 (2009) 679–692

Contents lists available at ScienceDirect

Journal of Mathematical Economics

journal homepage: www.elsevier.com/locate/jmateco

A duality theory of payment systems夽

Rodrigo Peñaloza
Center of Research in Economics and Finance (CIEF), University of Brasilia, 70910-900 Brasilia, Brazil

a r t i c l e i n f o a b s t r a c t

Article history: We model the Central Bank’s management of intraday liquidity in modern real-time gross
Received 23 February 2007 settlement systems as a linear programming problem parameterized by different intraday
Received in revised form 29 February 2008
monetary policies, such as reserve requirements, net debit caps and Lombard loans. We
Accepted 15 May 2008
then use duality theory to determine the shadow-prices of constraints of each bank. These
Available online 29 May 2008
shadow-prices can be used by the Central Bank to set personalized intraday monetary poli-
cies in order to reduce idleness of money and to give a microfoundation of the too-big-to-fail
JEL classification:
E58 policy.
© 2008 Elsevier B.V. All rights reserved.
Keywords:
Central banking
Real-time gross settlement systems
Shadow-prices

1. Introduction

A payment system is a network of banks in which funds transfers are made during the day. There are basically two types of
payment systems. In the Deferred Net Settlement (DNS) system, interbank payments are cleared, netted out and settled with
finality in the end of the day. Under the DNS system, banks clearly economize on liquidity needs, but are prone to systemic
risks. The increase of systemic risk in DNS systems due to the increasing value of interbank transfers has been a constant
concern for monetary authorities. The Bank for International Settlements (BIS) has therefore recommended the adoption of
real-time gross settlement (RTGS) systems for large-value transfers. In an RTGS system, interbank payments are settled, as
they are sent, by their gross amount. In other words, no bank can be illiquid at any given moment. This clearly reduces the
time lag between delivery of payment messages and final settlement, hence reducing systemic risk. However, the holding of
reserve money becomes a cost for banks. Indeed, since no illiquidity is allowed during the day, banks have to hold too much
liquidity for settlement purposes. Any miscalculation obliges the bank to obtain liquidity from other sources. The holding
of excess liquidity for settlement purposes is known to be the main problem of RTGS systems, for a huge amount of money
remains idle. We call it liquidity idleness. In order to facilitate the flow of payments and to reduce liquidity idleness, some
systems allow for queueing of payments, net debit caps, Lombard loans from the Central Bank and splitting of payments.
Net debit cap is a controlled overdraft that is granted to a bank. It allows the bank to become temporalily illiquid at some
cost. A Lombard loan is a loan from the Central Bank to a particular bank. Usually the Central Bank lends against collateral
and requires the bank to buy the collateral back in the end of the day. At the moment of lending, the price of the collateral
is slightly reduced. This procedure is called a haircut. Some countries allow for splitting of payments. Instead of paying all
or nothing, a bank can settle a fraction of the payment at one moment and settle the remaining fraction later in the day. We

夽 A prior version of this paper has circulated under the title “Shadow-prices in payment systems”, Discussion Paper #241, Department of Economics,
University of Brasilia.

0304-4068/$ – see front matter © 2008 Elsevier B.V. All rights reserved.
doi:10.1016/j.jmateco.2008.05.002
680 R. Peñaloza / Journal of Mathematical Economics 45 (2009) 679–692

call all these instruments intraday monetary policies. We refer to the RTGS report (BIS, 1997) for further details on payment
systems.
In this paper we solve the following problems: (a) How can the Central Bank set intraday monetary policies so as to minimize
liquidity idleness? (b) How can the Central Bank measure the marginal effect of a bank’s dollar withdrawal from the system on
the optimum flow of payments? Answering these questions is important. Problem (a) is a question that every central banker
asks and it deserves attention. Besides, knowing the marginal effect of a dollar withdrawal from the system gives the Central
Bank a sharp idea about the consequences of its intraday monetary policies. We can think of it as a measure of marginal
contagion.
In order to solve these problems we model the management of payment systems by the Central Bank as an infinite-
dimensional linear programming problem. Taking interbank payments as given as well as intraday monetary policies, the
Central Bank chooses how to settle payments in order to minimize liquidity idleness. We then use duality theory to find
the shadow-prices of liquidity constraints of each bank. The goal of our model is to use such shadow-prices to draw policy
implications. The overall conclusion is that intraday monetary policies should be personalized.
The basis for our main contributions arise from the very geometric and functional analitic framework of the model, which
is a novelty in the literature on payment systems. First, we show that interbank transfers and settlement of payments are
linked by a dual relationship. Then we prove the existence of an optimal way for the Central Bank to settle interbank payments.
Finally, we use duality theory to answer questions (a) and (b) above. As a by-product of our duality theory, we provide a
shadow-price argument for the personalized treatment of banks in payment systems, which is in line with the economic
wisdom that advocates that agents (in our case, banks) be priced according to their marginal values. Our duality theory of
payment systems shed some light on the t oo-big-to-fail policy in that big banks tend to have very high shadow-prices, which
explain their amplified impact on the system. Our model then offer a shadow-price argument in favor of that policy.
The infinite-dimensional linear programming framework has been widely used in general equilibrium analysis and other
allocational problems (see, for instance, Anderson and Nash, 1987; Gretsky et al., 1992, 1999, 2002; Makowski and Ostroy,
2000), but not in the payment systems literature. The construction of our model is entirely based on the BIS report on
real-time gross settlement systems (BIS, 1997).
The literature has focused on the influence of payment systems design on the behavior of banks (see DeBandt and
Hartmann, 2000 for a survey). Freixas and Parigi (1998) used the Diamond-Dybvig framework to make comparative wel-
fare analysis of DNS versus RTGS in a two-period two-island model with a liquidity shock and – the novelty of their model
– a locational shock. They showed that, if returns are stochastic, then an RTGS system dominates DNS system whenever
it is costly to keep an inefficient bank open, the probability of high return is low and the proportion of consumers that
have to consume in another location is low. In other words, RTGS system is preferred when the probability of a bank fail-
ure is high. Bech and Garrat (2001), in a very interesting model, use a Bayesian game to analyze the strategic behavior of
banks under the real-time gross settlement system. They show that intraday credit policies play an important role in the
banks’ decision to delay payments. Building on the literature on precautionary demand for reserves, Angelini (1998) shows
that priced intraday liquidity creates an incentive for banks to delay payments and that the outcome will not be socially
optimal.
Contrarywise to the papers cited above, we focus on the Central Bank itself. Some recent research has given atten-
tion to it (see, for instance, Angelini, 1998; Roberds, 1999) but our approach is quite different. The only literature whose
approach to problem (a) somehow resembles ours is the one on gridlock resolution. Leinonen and Soramäki (1999) con-
clude, from simulations, that an RTGS sytem with queueing is always more efficient than a net settlement system with
batch processing. Bech and Soramäki (2001) present a model of gridlock resolution, in which the objective is to maximize
the flow of outgoing queued payments subject to liquidity constraints faced by banks under RTGS systems. Building on
the bank clearing problem modelled by Güntzer et al. (1998), they introduce an additional constraint, called the sequence
constraint, which states that payments have to be settled in a predefined order, usually the FIFO (first-in-first-out) rule.
The general principles presented in Bech and Soramäki (2001) make their model the only one that is somehow related to
ours.
We conclude that there has been no unique framework for the payment system problem. In this paper we offer
a framework that is different from what has been done so far, so we hope to shed some light on the linear struc-
ture of payment systems and on the importance of shadow-prices as a powerful instrument in the hands of the central
banker.
The framework of the model is presented in Section 2. In Section 3 we state a theorem regarding the existence
of a solution to the Central Bank’s problem. The main contribution of our model lies however on the dual prob-
lem associated with the Central Bank’s liquidity management problem. The dual problem is presented in Section
4. The dual solution gives the shadow-prices of banks. These shadow-prices can be used by monetary authorities
to calculate the effect of personalized intraday monetary policies, such as reserve requirements, provision of Lom-
bard loans, net debit caps, etc. We show how shadow-prices can help determine intraday monetary policies so as
to bring systemic liquidity idleness down to zero. Therefore, reserve money can be fully used for settlement pur-
poses, avoiding waste of systemic liquidity. The only intraday monetary policy that we do not address here is the
queueing of payments. Elsewhere we include queueing and extend the model to another duality bracket (Peñaloza,
2002). Section 5 presents some policy implications. Section 6 concludes the paper. The proofs are presented in
Appendix A.
 R. Peñaloza / Journal of Mathematical Economics 45 (2009) 679–692 681

2. Real-time gross settlement systems

In this section we will present the set-up of the model. Our strategy is the following. First, we will define interbank
payments and the settlement of those payments as random functions over time belonging to a suitable dual pair of vector
spaces. Their inner product will be shown to be the total outflow of payments, which is just an affine transformation of
systemic liquidity idleness. The special features of the RTGS system, as described by the BIS (1997) report, will be given by
a set of linear inequality constraints determined by a linear operator. With these elements in hand, we will pose the Central
Bank’s problem of intraday liquidity management as a linear program.

2.1. Interbank payments

From a practical point of view, interbank payment messages arrive at the Central Bank at sporadic discrete times, so that
they can be regarded as a sequence. Since several payment messages may arrive on a minute-by-minute or even second-by-
second basis during the day, we will assume that time is continuous and that interbank payments are measurable functions
on the time interval.1
Let T = ([0, T ], B([0, T ]), ) be a measure space representing the time interval (a business day), where  is the Lebesgue
measure on the Borel- -algebra B([0, T ]). Recall that T is an atomless complete finite measure space.2 Let L∞ () be the space
of measurable bounded real functions on [0, T ], endowed with the sup-norm.
There is no way for the Central Bank to know ex ante what the pattern of payments will precisely be during the day. This
pattern is uncertain. Usually it is concentrated towards the end of the day, but, again, it is random (see McAndrews and Rajan,
2000). Therefore, the flow of payments is the sample path of a stochastic process. To model this feature, let (˝, , ) be a
complete atomless probability space.
Any participant of the payment system will be called a bank. B = {1, . . . , n} denotes the set of banks. Let xij (t, ω) be the
monetary value of the payment from bank i to bank j recorded on the books of the Central Bank at time t and state ω. Assume
that xij : [0, T ] × ˝ → [0, ∞) is a jointly measurable function such that xij (t, ·) ∈ L∞ ()+ , ∀t ∈ [0, T ], and, xij (·, ω) ∈ L∞ ()+ ,
 − a.e. Now define the function xi : [0, T ] → L∞ (, Rn )+ , by xi (t, ·) = (xij (t, ·))j ∈ B , where xij (t, ·) ∈ L∞ ()+ , ∀t ∈ [0, T ]. Finally,
define x = (xi )i ∈ B .
Given  = n2 , let E = L1 (, R ) be the space of integrable functions with values in R , endowed with the weak topology.
Recall that E is a locally convex, separable, metrizable linear topological space and that its topological dual is E ∗ = L∞ (, R ).
Whenever f ∈ E (or E ∗ ), we assume that f = (f1 , . . . , fn ) and fi = (fij )j ∈ B .
Thus, by our assumptions, for each t ∈ [0, T ], x(t) ∈ E ∗ = L∞ (, R ) and x ∈ L∞ (, E ∗ ). The space of payments is then the
positive cone of the space of all (equivalence classes of) Bochner-integrable functions3 that are essentially bounded, where
the norm is given by x = ess sup{x(t) : t ∈ [0, T ]}.

2.2. Settlement functions

When a bank sends an interbank payment message, the Central Bank immediately checks whether the sending bank has
sufficient funds. If it does, the payment is settled with finality. If it does not, what happens next depends on the design of
the payment system. If there is a centralized queueing facility, the message will be queued for later settlement. If there is no
queueing facility, then the message is returned. As we mentioned before, in this version of our model we assume that the
RTGS system has no centralized queueing facility, so the payment is either settled immediately with finality or is not. Such a
rule makes the constraint set nonconvex. We get rid of such nonconvexity by assuming fractioning of payments. By allowing
fractioning we also get an idea about how well the system could do. Besides, if a payment could be partially settled, then the
unsettled part would be obviously returned, in which case the sending bank would have to choose a better time to settle it.
The unsettled part of the payment would then be just another payment message sent later in the day. From the point of view
of the Central Bank, it makes no difference. The important thing is that fractioning of payments is clearly an improvement
over the all-or-nothing settlement rule.
We will now construct a definition of settlements function that is dual to the definition of interbank payments function.
Consider the following subset of E:

K = {q ∈ E : q(ω) ∈ [0, 1] ,  − a.e. and qii (ω) = 0, ∀ω ∈ ˝, ∀i ∈ B}.

Clearly, K ⊂ E = L1 (, R ) is convex and nonempty. An element of set K specifies the fractions of payments that must be
settled. Below, when we define what a settlement function is, such a function will take values on K. The meaning of qii (ω) = 0

1
Of course, measurability here is understood as Borel-measurability. Measurability is the weakest assumption we can make without loosing mathematical
structure.
2
Notice that  may also be any atomless measure other than Lebesgue’s. 
3
Given a finite measure space (, г, ) and a Banach space E, a -measurable function f :  → E is Bochner-integrable if, and only if, f d < ∞

(Diestel and Uhl, 1977).
682 R. Peñaloza / Journal of Mathematical Economics 45 (2009) 679–692

is that a bank cannot settle a payment it sends to itself, because there is no such thing as a self-transfer. We need the following
assumption:

Assumption A1. The set K is weakly compact separable and metrizable for the relative topology induced from L1 (, R ),
where the metric on K is denoted by m, and uniformly integrable.

Let i : [0, T ] → L1 (, [0, 1]n ), be defined by the array i (t) = ( ij (t))j ∈ B , where ij (t) ∈ L1 (, [0, 1]), ∀t ∈ [0, T ]. Let  = n2 .
n
For ease of notation, we will use the isometric isomorphism L1 (, [0, 1]n ) ≈ L1 (, [0, 1] ). Assume that : [0, T ] → K is
(B([0, T ]), B(K))-measurable, where B(K) is the Borel- -algebra generated by the relative weak topology on K inherited from
E. We call a settlement function. Let L1 (, K) be the set of Bochner-integrable settlement functions. We have that, for each
t ∈ [0, T ], (t) ∈ E = L1 (, R ) and ∈ L1 (, E). To understand what means, pick a time t. Then (t) is a random vector. If
the state of the world is ω, then (t)(ω), wich can be written as (t, ω), is given by an array ( ij (t, ω))i,j=1,...,n whose typical
element ij (t, ω) specifies the fraction of the payment xij (t, ω) from bank i to bank j at time t and state ω that will be settled
with finality. If the payment system allows no fractioning, then (t, ω) will be a vector of 0’s and 1’s. Otherwise, it will be a
vector of numbers between 0 and 1.

Definition 1. A settlement function is any Bochner-integrable function : [0, T ] → K, that is, any Bochner-integrable
function ∈ L1 (, E) given by (t)(ω) = (t, ω) = ( 1 (t, ω), . . . , n (t, ω)) such that ij (t, ω) ∈ [0, 1], ∀i, j ∈ B, ii (t, ω) ≡ 0,
∀i ∈ B, ∀t ∈ [0, T ],  − a.e.

2.3. Duality between payments and settlements

Settlement functions and interbank payments functions are then dual concepts. To be more specific, on the dual system
L1 (, E), L∞ (, E ∗ ) consider the bilinear form · : L∞ (, E ∗ ) × L1 (, E) → R given by
 T  
x· = xij (t, ω) ij (t, ω) d(ω) d(t)
0 ˝
i∈B j∈B

Now it is easy to see the topological duality between payments and settlements. Given the Banach space E = L1 (, R ) and
its dual E ∗ = L∞ (, R ), we have:

1. ∈ L1 (, E) denotes the settlement function, with the condition that (t) ∈ K, ∀t ∈ [0, T ]. The space of settlement functions
is endowed with the weak topology.
∗
2. x ∈ L∞ (, E ∗ ) = (L1 (, E)) denotes the payments function, which belongs to the dual space, endowed with the weak∗
topology.
3. L1 (, E), L∞ (, E ∗ ) is the topological dual pair denoting the duality between interbank payments and settlements with
duality bracket representing the expected total outflow of interbank payments, i.e., x · .

2.4. Feasibility

In this section we introduce a notion of feasibility for settlement functions in an RTGS system. The idea behind feasibility
is that, at any moment, the total value of settled outgoing payments cannot exceed the individual bank’s resources. Resources
refer to sources of intraday funds, which include initial balance held at the bank’s Central Bank account, net transfers received
from other banks up to the time of current settlements, and intraday credit extensions from the Central Bank, such as
repurchase agreements, collateralized loans and/or net debit caps.
Consider the set-valued map
: [0, T ] × K  K given by4 :
⎡ ⎤
   

(t, q) = { ∈ K : xij (t, ω)ij (ω) ≤ B0i + ⎣ xji (, ω)qji (ω) − xij (, ω)qij (ω)⎦ d() + D̄i (t, ω)
[0,t)
j∈B j∈B j∈B

+zi (t, ω),  − a.e., ∀t ∈ [0, T ], ∀i ∈ B}

Here, B0i is the initial balance of bank i at its Central Bank account. It is the minimum balance that the Central Bank requires
bank i to hold in the beginning of every day.
The function D̄i denotes the net debit cap for bank i. It is a source of liquidity offered by the Central Bank to bank i. The
overall effect of a net debit cap is that bank i is allowed to incur temporary intraday overdrafts up to a certain amount. Let
D̄ : [0, T ] × ˝ → [0, ∞)n be defined by D̄(t, ω) = (D̄1 (t, ω), . . . , D̄n (t, ω)), where D̄i (t, ω) is a net debit cap for bank i at time

4
Given two nonempty sets, X and Y, we prefer to use the notation
: X  Y for a correspondence instead of the more cumbersome notation
: X → 2Y ,
where 2Y is the set of subsets of Y.
 R. Peñaloza / Journal of Mathematical Economics 45 (2009) 679–692 683

t and in state ω. Though a net debit cap contradicts the very philosophy of real-time gross settlement systems, it is still an
available intraday monetary policy.
Finally, let zij (t, ω) > 0 be the amount of the collateralized loan received by bank i from the Central Bank at time t and
state ω in order to fullfil its payment obligations toward bank j. Whenever zij (t, ω) < 0, we say that bank i is paying back
T
a loan. Define zi (t, ω) = z (t, ω).
j ∈ B ij
We require that 0
zi (t, ω) d(t) = 0,  − a.e., that is, by the end of the day, every
loan is payed back. If xij (t, ω) = 0, then zij (t, ω) = 0.5 According to the above condition, no positive repo price is charged over
collateralized loans. We could nevertheless easily introduce positive repo prices into the model. Partition the time interval into
three disjoint Borel-measurable sets, say, [0, T ] = Ti0 ∪ Ti1 ∪ Ti2 , where Ti0 , Ti1 , Ti2 ∈ B([0, T ]), and Ti0 ∩ Ti1 , Ti0 ∩ Ti2 , Ti1 ∩ Ti2 = ∅.
No repo loan is made or payed back at any t ∈ Ti0 , that is, zi (t, ω) = 0, ∀t ∈ Ti0 ,  − a.e. Loans are received at any time t ∈ Ti1 ,
i.e., zi (t, ω) > 0, ∀t ∈ Ti1 ,  − a.e., whereas loans are payed back at any t ∈ Ti2 , i.e., zi (t, ω) < 0, ∀t ∈ Ti2 ,  − a.e. We would
  T
then require that Ti1
zi (t, ω) d(t) ≤ i zi (t, ω) d(t),  − a.e. This implies that zi (t, ω) d(t) ≤ 0,  − a.e. A repo price for
T2  0
bank i is a function ri (t, ω) satisfying Ti1
zi (t, ω)(1 + ri (t, ω)) d(t) = Ti2
zi (t, ω) d(t),  − a.e. For simplicity however we will
T
assume that no repo price is charged, i.e., 0 zi (t, ω) d(t) = 0,  − a.e. In any case, let z : [0, T ] × ˝ → [0, ∞)n be defined by
z(t, ω) = (z1 (t, ω), . . . , zn (t, ω)).
Let = (B0 , D̄, z) be the parameter profile and assume that there is a compact set ˘ ⊂ [0, ∞)n × L∞ (, E ∗ ) such that
2

∈ ˘. In order to make the role of the parameters explicit, we may write the constraint correspondence
as
. For
simplicity, define gi (t, ω) = B0i + D̄i (t, ω) + zi (t, ω), and let g(t, ω) = (g1 (t, ω), . . . , gn (t, ω)). Of course the function gi inherits
all the restrictions imposed on D̄i and zi . All these will be part of set ˘. The terms in the definition of g will be called intraday
monetary policies. In other words, an intraday monetary policy is any parameter function that can be set discretionarily by
the Central Bank in order to smooth out the flow of payments during the business day.
The set-valued map
represents the constraint set that any bank has to satisfy in an RTGS system. Before defining
feasibility, we introduce an auxiliary notion of feasibility, which we call q-feasibility.

Definition 2. A settlement function ∈ L1 (, K) is q-feasible if, given q ∈ K, is a Bochner-integrable selection of the
set-valued map
(·, q) : [0, T ]  K, i.e., (t) ∈
(t, q), ∀t ∈ [0, T ].

The interpretation of q-feasibility is straightforward. At any given time t, a settlement function is q-feasible if it can settle
some of the payments at time t given that the settlement function q is the

 settlement been in place upto time
function that had
t. If (t) ∈
(t, q), then (t) ∈ K and x (t, ω) ij (ω) ≤ gi (t, ω) + [0,t)
j ∈ B ij
x (, ω)qji (ω) −
j ∈ B ji
x (, ω)qij (ω) d().
j ∈ B ij
The left-hand side is the sum total of all payments bank i makes at time t and state ω if the settlement function at that
time and that state is (evaluated at t). The right-hand side is the total funds bank i gets from the Central Bank in the
form of intraday monetary policies plus net transfers received from every other bank up to that point in time, given that the
settlement function adopted by the system from the beginning up to time t was q.
In all payment systems in the world, the settlement of payments obey a constant rule (e.g. no queueing) previously
agreed upon between the Central Bank and participating banks. This is a legal issue that cannot be ruled out of any model of
the payment system. The Bank for International Settlements has always highlighted the importance of legal issues and the
necessity of having clear rules for the payment systems (BIS, 1997, p. 30). This important legal feature of payment systems
will be incorporated in our constraint set by a fixed point argument. The settlement function ∈ L1 (, K) is feasible if is
-feasible, i.e., (t) ∈
(t, (t)), ∀t ∈ [0, T ].
To see that this is the notion of feasibility that makes sense in our context, notice that, if (t) ∈
(t, (t)), ∀t ∈ [0, T ], then
is -feasible and therefore the feasibility constraint of the RTGS system:
⎡ ⎤
   
xij (t, ω) ij (t, ω) ≤ B0i + ⎣ xji (, ω) ji (t, ω) − xij (, ω) ij (t, ω)⎦ d() + D̄i (t, ω) + zi (t, ω),  − a.e.,∀i ∈ B
[0,t)
j∈B j∈B j∈B

is satisfied, ∀t ∈ [0, T ]. At time t, the settlement function applied to current payments (the sum in left-hand side of the
constraint) is the same one that had been applied to all payments settled up to time t (the sum in the right-hand side of the
constraint).
The set FIX(
) = {f ∈ L1 (, K) : f (t) ∈
(t, f (t)), ∀t ∈ [0, T ]} is called the set of random fixed points 6 o f
. Then a settle-
ment function is f easible if it is a random fixed point of the constraint set-valued map
.

Definition 3. Let ∈ L1 (, K) be a settlement function. Consider the set-valued map

: [0, T ] × K  K given above and
let FIX(
) be the set of random fixed points of
. We say that is feasible if ∈ FIX(
), i.e., if (t) ∈
(t, (t)), ∀t ∈ [0, T ].

5
I thank a referree for pointing out the necessity of this condition.
6
The term random fixed point stands for the fact that in the mathematical literature, T is usually a probability space, not a time space. It is convenient
however to stick with the standard terminology.
684 R. Peñaloza / Journal of Mathematical Economics 45 (2009) 679–692

Definition 4. A real-time gross settlement system is a collection

 = {L (, E), L∞ (, E ∗ ) , (

) ∈ ˘ },
1

where L1 (, E), L∞ (, E ∗ ) is the duality pairing of settlements and payments,
: [0, T ] × K  K is the set-valued random
operator of RTGS constraints, and ˘ is the set of intraday monetary policies.

2.5. The Central Bank’s liquidity problem

In this section we formulate the problem to be solved by the Central Bank.

First, we will define the objective function step-by-step. According to the BIS, the net intraday liquidity at certain time t is
given by total initial balances (the sum of B0i over i ∈ B) held on the Central Bank accounts minus the total value of outgoing

payments L(t) = (Bi −
i∈B 0
x (t)), where xij (t) denotes the payment from bank i to bank j at time t. This definition
j ∈ B ij
is too pessimistic in that it overestimates actual liquidity needs. In order to account for the role
of fractional
settlements
and to be more carefull about actual liquidity needs, we make a slight modification7 : L(t) = (B i − x (t) ij (t)).
i∈B 0 j ∈ B ij
Minimizing liquidity needs according to this definition would require minimization at every point in time. In order to get
rid of this extremely costly procedure, we modify it even further by considering the sum over t or the integral over t of L(t),
T
so that ( ) = (1/T ) 0
L(t) d(t). Since we also want to introduce randomness into the system, we came to the following
definition of aggregate liquidity, which we call systemic liquidity idleness 8 :
  T  
1
( ) = B0i − xij (t, ω) ij (t, ω) d(ω) d(t)
T 0 ˝
i∈B i∈B j∈B

Our definition is better for two reasons. First, it accounts for the history of payments within the day. Second, what really
matters is the total flow of payments in the end of the day, which is accounted for by the summation. The standard definition,
on the contrary, requires optimization at every point in time. Suppose that ( ) =1 billion. This means that there are $1
billion in the system that are not being used for settlement purposes. The reason could be that the settlement rule, for
example, is of the 0 − 1 type, no fractioning being allowed. By introducing fractioning, the enlargement of the feasible set
would reduce the value of ( ) and make the payment system economize on liquidity.
Since initial balances held at Central Bank accounts are fixed (as well as T), minimizing systemic liquidity idleness is
equivalent to maximizing total expected value of the flow of outgoing payments (the second term in the definition of (·)).
In other words, the Central Bank wants to solve:
⎧  T  
⎪
⎪
⎪
⎪ sup xij (t, ω) ij (t, ω) d(ω) d(t)
⎪
⎪ 1
⎨ ∈ L (,K) 0 ˝ i ∈ B j ∈ B ⎡ ⎤
   
⎪
⎪ s.t. xij (t, ω) ij (t, ω) ≤ gi (t, ω) + ⎣ xji (, ω) ji (t, ω) − xij (, ω) ij (t, ω)⎦
⎪
⎪
⎪
⎪ [0,t)
⎩ j ∈ B j ∈ B j ∈ B
× d(t)∀i ∈ B, ∀t ∈ [0, T ],  − a.e.

Yet another simpler notation is:


sup x·
(P) ∈ L1 (,K)
s.t ∈ FIX(
)

A solution to problem (P) above will be called an optimal settlement function.

The left-hand side of the constraint is the total value of outgoing payments from bank i at time t and state ω. The second
term on right-hand side is the net transfer up to time t to bank i’s account. The first term refers to any intraday monetary
policy specific to bank i. The objective function is total outflow of payments.

3. Existence of optimal settlement functions

In order to show that an optimal settlement function exists, we need to assume the following:

Assumption A2. ∀q ∈ K,
(·, q) : [0, T ]  K is (B([0, T ]), ℘)-measurable, where ℘ is the Borel- -algebra generated by the
weak topology on K. Moreover, it is closed-valued and, ∀t ∈ [0, T ],
(t, ·) : K  K is nonexpansive, i.e., H(
(t, p),
(t, q)) ≤
m(p, q), ∀p, q ∈ K, where H is the Hausdorff metric on 2K induced by the metric m on K.

7
With some abuse of notation, we used the same letter, L, but we hope this causes no confusion.
8
I thank Joe Ostroy for naming the function  as liquidity iddleness.
 R. Peñaloza / Journal of Mathematical Economics 45 (2009) 679–692 685

Our main theorem in this section states that problem (P) has a solution, that is, there exists an optimal setllement function
that minimizes liquidity idleness.

Theorem. Assume (A1) and (A2). Then there exists an optimal settlement function, ∗ , that solves (P), i.e., arg sup{x · :
/ ∅. Moreover, ∗ satisfies x ∈ NFIX(
) ( ∗ ), where NFIX(
) ( ) is the normal cone to FIX(
) at ∗ .
∈ FIX(
)} =

The proofs have been relegated to Appendix A.

4. Shadow-prices of intraday monetary policies

Having set the Central Bank’s problem as a linear program, it is straightforward to write down the dual linear program.
To make things easier, we have however to rewrite the primal problem in matrix form. This is what we will do first. Standard
duality theory will give us the dual problem in matrix form. We then have to make explicit the typical dual constraints. The
solution to the dual problem is the collection of Lagrange multipliers of the constraints. The importance of writting down the
dual problem is that the Lagrange multipliers are the shadow-prices of the intraday monetary policies. Finally, since there
will be no duality gap, we can get an equation that describes nil liquidity idleness in terms of the shadow-prices of intraday
monetary policies. Such an equation will be crucial for us to draw the policy implications in the next section.
Define, ∀i ∈ B, ∀(t, ω) ∈ [0, T ] × ˝, the following matrices:
⎡ 0 ··· 0
⎤
⎢ .
.. .. .. ⎥
⎢ . . ⎥
⎢ 0 ··· ⎥
⎢ 0 ⎥
Xi (t, ω) = ⎢
⎢ xi1 (t, ω) · · · xin (t, ω) ⎥
⎥ ← ith row
⎢ 0 ··· 0 ⎥
⎢ ⎥
⎣ .. .. .. ⎦
. . .
0 ··· 0 n×n

⎡ ⎤
xi1 (t, ω) 0 ··· 0
⎢ 0 xi2 (t, ω) ··· 0 ⎥
Yi (t, ω) = ⎢ .. .. .. ⎥
⎣ . .
..
. .
⎦
0 0 ··· xin (t, ω) n×n

Consider the partitioned matrices:

X(t, ω) = [X1 (t, ω) | · · · | Xn (t, ω)]n×n2

Y (t, ω) = [Y1 (t, ω) | · · · | Yn (t, ω)]n×n2

Then the expression x (, ω) ji (, ω) −
j ∈ B ji
x (, ω) ij (, ω),
j ∈ B ij
∀i ∈ B, can be written in matrix form as

Y (, ω) (, ω) − X(, ω) (, ω) = [Y (, ω) − X(, ω)] (, ω), ∀ ∈ T,  − a.e.

In order to write down the dual problem properly, the constraint 0 ≤ ij (t, ω) ≤ 1, ∀t ∈ [0, T ],  − a.e., ∀i, j ∈ B, must be
written explicitly in the feasible set. Relax the condition ∈ L1 (, K) to ∈ L1+ (, E) with the added constraint ≤ 1, where
1 is the vector with entries equal to 1. We call it the c onsistency constraint. Therefore we have:
⎧  T
⎪
⎪
⎪
⎪ sup x(t, ω) (t, ω) d(ω) d(t)
⎨ ∈ L1 (,E) 0 ˝
 t
⎪
⎪ s.t. X(t, ω) (t, ω) ≤ g(t, ω) + [Y (, ω) − X(, ω)] (, ω) d()
⎪
⎪
⎩ 0
(t, ω) ≤ 1, ∀t ∈ T,  − a.e.
t
The constraint X(t, ω) (t, ω) ≤ g(t, ω) + 0 [Y (, ω) − X(, ω)] (, ω) d() is the set of all feasibility constraints in the
RTGS system. The parameters of these constraints are summarized by the map g.
Define the operator ˚ : L1 (, E) → L∞ (, E) by
 t
˚( )(t, ω) = X(t, ω) (t, ω) − [Y (, ω) − X(, ω)] (, ω) d()
0
686 R. Peñaloza / Journal of Mathematical Economics 45 (2009) 679–692

and let the operator  : L∞ (, E ∗ ) → L1 (, E ∗ ) be given by

 T
 ()(t, ω) = (t, ω)X(t, ω) − (, ω)[Y (, ω) − X(, ω)] d()
t

The following proposition shows that  is the adjoint of ˚. We need it to write down the dual feasible set. The proof has
been relegated to Appendix A.

Proposition. The operators ˚ and  satisfy , ˚ = , , hence  = ˚∗ , i.e.,  is the adjoint of ˚.

Thus the paired linear programming problems in an RTGS system are given by
⎧
⎪ sup x· ⎧
⎪
⎨ ⎨ (,)
inf ·g+·1
Primal s.t. ˚ ≤ g Dual ∗ ∗
⎪
⎪ I ≤ 1 ⎩ s.t. ˚  + I  ≥ x
⎩ ,  ≥ 0
≥0

In other words, the dual problem is

⎧  T  T
⎪
⎪ (t, ω)g(t, ω) d(ω) d(t) + 1 · (t, ω) d(ω) d(t)
⎪
⎪ inf
⎪
⎨ (,) 0 ˝  T0 ˝
s.t. (t, ω)X(t, ω) + (t, ω) ≥ x(t, ω) + (, ω)[Y (t, ω) − X(t, ω)] d()
⎪
⎪
⎪
⎪ t
⎪
⎩ (t, ω), (t, ω) ≥ 0, ∀t ∈ [0, T ],  − a.e.
(, ) ∈ L∞ (, K) × L∞ ()

Writting it down explicitly, the dual problem is

⎧  T   T  
⎪
⎪
⎪ (,)
⎪ inf i (t, ω)gi (t, ω) d(ω) d(t) + ij (t, ω) d(ω) d(t)
⎪
⎪
⎨ 0 ˝
i∈B
0

˝
i∈B j∈B
T

⎪
⎪ s.t. xij (t, ω)i (t, ω) + ij (t, ω) ≥ xij (t, ω) + xij (t, ω) (j (, ω) − i (, ω)) d()
⎪
⎪
⎪
⎪ i (t, ω), ij (t, ω) ≥ 0, ∀i, j ∈ B, ∀t ∈ [0, T ],  − a.e.
t
⎩
(, ) ∈ L∞ (, K) × L∞ ()

Here i (t, ω) is the Lagrange multiplier of the feasibility constraint of bank i at time t ∈ [0, T ] and state ω ∈ ˝. Thus i (t, ω)
is the shadow-price of the parameter gi (t, ω) and captures the marginal effect over the maximum flow of payments of a dollar
increase of the intraday monetary policy toward bank i. On the other hand, ij (t, ω) is the shadow-price of the consistency
constraint 0 ≤ ij (t, ω) ≤ 1.
Notice that, even though the space of dual programs is L1 (, K) × L1 (), we want to find solutions in the linear subspace
L∞ (, K) × L∞ (). The boundedness and measurability of x and g guarantee that strong duality holds (Tyndall, 1967), so the
primal and dual values coincide. Plugging the dual value into the definition of systemic liquidity idleness, we get:
  T  
1
= B0i − xij (t, ω) ij (t, ω) d(ω) d(t)
T 0 ˝
i∈B i∈B j∈B

  T    T  
1 1
= B0i − i (t, ω)gi (t, ω) d(ω) d(t) − ij (t, ω) d(ω) d(t)
T 0 ˝
T 0 ˝
i∈B i∈B i∈B j∈B

Define the following expected averages of shadow-prices over time:

 T 
1
¯ ie = i (t, ω) d(ω) d(t)
T
 0T  ˝
1
¯ ije = ij (t, ω) d(ω) d(t)
T 0 ˝

and recall that gi (t, ω) = B0i + D̄i (t, ω) + zi (t, ω). Then a simple algebraic manipulation shows that:

   T  
1
= B0i (1 − ¯ ie ) − ¯ ije − i (t, ω)(D̄i (t, ω) + zi (t, ω)) d(ω) d(t)
T 0 ˝
i∈B i∈B j∈B i∈B
 R. Peñaloza / Journal of Mathematical Economics 45 (2009) 679–692 687

The Central Bank wants to use shadow-prices to set the control variables (initial balances, net debit caps, and Lombard
loans) so as to bring systemic liquidity idleness down to its minimum level,  = 0. When systemic liquidity idleness is nil,
the design of the payment system allows interbank payments to flow smoothly without any money being unused.

Definition 5. A systemic monetary policy ∈ ˘ is liquidity-efficient if systemic liquidity idleness is zero,  = 0.

Then the fundamental equation for liquidity-efficiency becomes:

   T  
1
B0i (1 − ¯ ie ) − ¯ ije = i (t, ω)(D̄i (t, ω) + zi (t, ω)) d(ω) d(t)
T 0 ˝
i∈B i∈B j∈B i∈B

As long as intraday monetary policies are set according to the fundamental equation for liquidity-efficiency, there is no waste
of liquidity, hence no deadweight losses for the banking sector.

5. Policy implications

In this section we present some policy implications drawn from our dual approach. To illustrate how our model can
answer questions (a) and (b), suppose systemic liquidity idleness is positive,  > 0, which means that some reserve money
is being unused.
How can the Central Bank bring systemic liquidity idleness down to zero? Decreasing reserve requirements uniformly
across banks is not the answer, though it is what Central Banks usually do. Indeed, the Central Bank should increase reserve
requirements of those banks with average expected liquidity shadow-prices above unity, (1 − ¯ ie ) < 0. Some banks may have
their reserve requirements increased by as much as everybody else, but this extra reserve money may become stuck, for
what these banks had were already enough for their settlement purposes. Thus any extra money becomes unused, and its
opportunity cost represents an economic loss.
Suppose, for instance, that the only source of intraday liquidity, besides net transfers, is the amount of initial reserves. If
systemic liquidity is positive, then:
 
0<= B0i (1 − ¯ ie ) − ¯ ije
i∈B i∈B j∈B

How could the Central Bank have set reserve requirements in the first place so as to reduce systemic liquidity idleness in
the RTGS system? In the simple case we are analyzing, the answer is given by the equation:
 
B0i∗ (1 − ¯ ie ) − ¯ ije = 0
i∈B i∈B j∈B

One possible solution to it is

⎧  
⎪
⎪ {B0k (1 − ¯ ke ) − ¯ kj
e
} ¯ ije
⎨
k ∈ B0 j∈B j∈B
B0i∗ = − , if i ∈ B1
⎪
⎪ #B1 (¯ ie − 1) ¯ ie − 1
⎩
B0i , if i ∈ B0

Here, B1 = {i : ¯ ie > 1} is the set of banks with expected average liquidity shadow-price strictly above unity, #B1 is its cardi-
nality, and B0 = {i : 0 ≤ ¯ ie ≤ 1} = B \ B1 is the rest of banks. The above solution says that banks with low shadow-prices are
required to keep their historical reserve balance, but banks with high shadow-prices are required to change initial reserves
to B0i∗ .
Among other things, the Central Bank can estimate the effect of different intraday credit policies. For example, what
happens when some bank is allowed to use net debit caps? Suppose the Central Bank could reward some banks with net
debit caps. Shadow-prices tell us the extent of such net debit caps. Suppose, for instance, that there are no Lombard loans
and that net debit caps are constant and uniform, i.e., banks get the same net debit cap, say D̄i (t, ω) = D > 0. If  > 0, what
would have been an optimal D? If the Central Bank introduces a net debit cap, then it wants to set D such that:
  
B0i (1 − ¯ ie ) − ¯ ije − D ¯ ie = 0
i∈B i∈B j∈B i∈B

that is:
 
B0i (1 − ¯ ie ) − ¯ ije
i∈B i∈B j∈B
D∗ = 
¯ ie
i∈B
688 R. Peñaloza / Journal of Mathematical Economics 45 (2009) 679–692

Net debit caps can be personalized. Indeed, if we go further into our example, suppose we want to set an optimal Di∗ . A
possible solution is obtained by setting:

B0i (1 − ¯ ie ) − ¯ ije
j∈B
Di∗ =
¯ ie

Notice that some banks will get net debit caps, which amount to a form of subsidy from the Central Bank, whereas other
banks will have to pay taxes. Indeed, depending on the shadow-prices, Di∗ may be positive or negative. Thus net debit caps
should be financed by banks themselves through redistribution of liquidity. While some banks get net debit caps, other banks
get a positive lower bound on current balance during the day.
From a political point of view, it is better to give banks with low shadow-prices no debit caps and to set another level of
net debit caps to banks with high shadow-prices. Of course, banks with high shadow-prices will have to fully bear the costs
of other banks not paying taxes. Again, consider the set of banks with expected average liquidity shadow-price strictly above
unity, B1 = {i : ¯ ie > 1}, and let B0 = {i : 0 ≤ ¯ ie ≤ 1} = B \ B1 denote the rest of banks. Then another solution is
⎧   
⎪
⎪ B0i (1 − ¯ ie ) − ¯ ije {B0k (1 − ¯ ke ) − ¯ kj
e
}
⎨
k ∈ B0 j∈B
Di∗∗ = j∈B
+ , if i ∈ B1
⎪
⎪ ¯ ie #B1 ¯ ie
⎩
0, if i ∈ B0

Comparing this result with the previous one, we see that net debit caps are reduced by the amount given by the second
term. The amount of reduction is exactly the amount of taxes not paid by banks with low shadow-prices divided equally
among banks with high shadow-prices and weighed by the inverse of expected average liquidity shadow-prices. Thus banks
with even higher shadow-prices are rewarded with a lower reduction of net debit caps, whereas banks with not so high
shadow-prices are punished with a higher reduction of net debit caps.
From duality theory we obtain a measure of marginal systemic effect (or a measure of marginal contagion) of an individual
bank on the overall systemic liquidity. Suppose again that the only source of intraday liquidity, besides interbank transfers,
is the level of initial reserves. Then:

MSEi = (1 − ¯ ie )B0i − ¯ ije
j∈B

is the sum of all terms of  due to the presence of bank i in the system. Suppose bank i fails within the business day
for whatever reason. Then MSEi is a monetary measure of the marginal effect of each dollar withdrawn from the system.
Notice that marginal contagion is decomposed into two terms. The first term is, to a first order approximation, the effect on
systemic liquidity that comes from the variation of bank i’s own liquidity. The second term is, to a first order approximation,
the network effect on systemic liquidity through interbank payments. We call MSEi the bank i’s measure of marginal contagion.
If, more generally, we consider the other intraday monetary policies, then marginal contagion is given by
  T 
1
MSEi = B0i (1 − ¯ ie ) − ¯ ije − i (t, ω)(D̄i (t, ω) + zi (t, ω)) d(ω) d(t)
T 0 ˝
j∈B

In the case there is no Lombard loan and net debit caps are constant but not uniform across banks, we have:

MSEi = B0i (1 − ¯ ie ) − ¯ ije − ¯ ie D̄i
j∈B

Thus marginal contagion is decomposed into three terms. The first two terms have already been interpreted above. The third
term is, to a first order approximation, the effect on systemic liquidity that comes from the allowance of overdraft as a source
of liquidity.
Another policy implication regards the too-big-to-fail policy. Few big banks are responsible for the majority of interbank
transactions and for most of the total amount of money that circulates in the payment system. Since the arising of central
banks in the nineteenth century it has been common sense that the failure of a big bank could be very pernicious to the
system. The main argument is that a very likely domino effect would harm other banks as well. The payment system is
an obvious conduit for the spreading out of financial problems. Large scale failures would then have a profound negative
impact on the real economy. Even though this issue has not been addressed here, our model can however shed some light
on it too. Big banks work with high initial balances and makes too many transactions. They have a more balanced inflow
and outflow of money transfers. They are also usually much more involved in high FX transactions than small banks (see
BIS, 1997 on RTGS systems, pp. 19–21). Liquidity constraints will very likely be biding and the flow of payments will be
very sensitive to small variations, because of the greater number of interconnections with other banks. Therefore liquid-
ity shadow-prices will be positive and very high. If, for example, ¯ ie = 10, then the maximum flow of payments will be
 R. Peñaloza / Journal of Mathematical Economics 45 (2009) 679–692 689

reduced by at least ten times the bank’s initial balance. Even a number as low as two, depending on the size of all other
banks in relation to the failing bank, will cause a harmful reduction of the maximum flow. Therefore shadow-prices too
can provide a solid microeconomic foundation for the too-big-to-fail policy. The argument in terms of shadow-prices has
not, at least to our knowledge,
been proposed yet in the literature. Notice that the fundamental equation for liquidity-
efficiency can be written as MSE = 0. If 
¯ e is too big, then MSE might be negative and with high absolute value.
i∈B i h h
If bank h fails and is removed from the system, liquidity-efficiency requires that MSEi increase for every other bank i. In
other words, everybody else will be in need of liquidity. The lower ¯ ie , the higher is the amount of extra liquidity bank i will
need.

6. Conclusion

We constructed a model of real-time gross settlement systems in which interbank payments and settlement rules are
linked by a dual relationship represented by the bracket duality between the space of payments and the space of settlements.
In spite of its apparent abstract nature, the constraint set is in accord with the actual liquidity constraints faced by banks in
RTGS systems. We believe we made a point in favor of microeconomic rationality in the management of intraday monetary
policies and in favor of the idea that banks should be treated according to their marginal values to the system. The policy
implications we drew showed us how the Central Bank could make this idea effective. In addition, we argued that our model
provides a microfoundation for the too-big-to-fail policy in terms of shadow-prices only.
From a practical perspective, our model could be used in the following way. A finite-dimensional version of it (see, for
instance, Peñaloza, 2003, 2005) could be applied to actual data on interbank transfers in an RTGS system for a period of time.
The Central Bank would then use these data to calculate shadow-prices of banks (their marginal systemic effects). Having
collected a time series of shadow-prices of each bank, the Central Bank could check for any pattern that might amelioriate
its policy decisions. Even if the Central Bank does not adopt personal treatment of banks in the way we advocate, it could
still use the collected shadow-prices to identify the banks that have more impact on the system.
A possible extension of our model is to include many future business days. Banks carry debts from one day to another.
Our model would then provide the shadow-prices of overnight loans.
Duality theory proved to be very useful in other areas of economic theory and we hope it somehow be relevant for the
microeconomics of payment systems as well.

Acknowledgements

I am deeply indebted to Joe Ostroy for his profound insights on this subject. I also thank David Levine, Alberto Bennardo,
Hans Scholhammer, and David Rahman for useful comments on the very first version of the paper and a referree for comments
on the final version.

Appendix A. Proofs

In this appendix we prove the lemmas and the main results. We need the following:

Definition. Let K ⊂ L1 (, R ) satisfy (A1). A set-valued map ˚ : K  K is demiclosed at 0 if, given qn →w q (i.e., weakly) and
pn ∈ ˚(qn ) such that pn → 0 (i.e., strongly), then 0 ∈ ˚(q).

The next series of lemmas will enable us to show that the constraint set is well-behaved and that the problem is well
posed.

Lemma 7. Assume (A1). Then, ∀t ∈ [0, T ], set-valued map Id(·) −

(t, ·) : K  K (where Id is the identity operator on K) is
demiclosed at 0, that is, if qn →w q (i.e., weakly) and pn ∈ Id(qn ) −
(t, qn ) is such that pn → 0 (i.e., strongly), then 0 ∈ Id(q) −
(t, q).

Proof of Lemma 1. Fix t ∈ [0, T ]. Define ˚t : K  K by  ˚t (q) = q −

(t, q). Let qn →w q weakly and take pn ∈ ˚t (qn ) = qn −

(t, qn ) such that pn → 0 (i.e., strongly). Then limn→∞ ˝ x ⊗ qn d = ˝ x ⊗ q d, ∀x ∈ L∞ (, R ). In particular:

 
xij (t, ·)qnij (·)→w xij (t, ·)qij (·)
j∈B

j∈B  
xji (, ·)qnji (·) d()→w xji (, ·)qji (·) d()
[0,t) [0,t)
 j∈B
  j∈B

xij (, ·)qnij (·) d()→w xij (, ·)qij (·) d()
[0,t) [0,t)
j∈B j∈B
690 R. Peñaloza / Journal of Mathematical Economics 45 (2009) 679–692

Since pn ∈ ˚t (qn ), ∃n ∈
(t, qn ) such that pn = qn − n , where:
⎡ ⎤
   
xij (t, ω)ijn (ω) ≤ B0i + ⎣ xji (, ω)qnji (ω) − xij (, ω)qnij (ω)⎦ d() + D̄i (t, ω) + zi (t, ω),
[0,t)
j∈B j∈B j∈B

 − a.e.,∀t ∈ [0, T ], ∀i ∈ B
 
and pn L1 = ˝
|pn | d = ˝
|qn − n | d → 0. By definition, n = qn − pn , so:
⎡ ⎤
   
xij (t, ω)(qnij (ω) − pnij (ω)) ≤ B0i + ⎣ xji (, ω)qnji (ω) − xij (, ω)qnij (ω)⎦ d() + D̄i (t, ω) + zi (t, ω),
[0,t)
j∈B j∈B j∈B

 − a.e.,∀t ∈ [0, T ], ∀i ∈ B

Thus:
⎡ ⎤
    
xij (t, ω)qnij (ω) − xij (t, ω)pnij (ω) ≤ B0i + ⎣ xji (, ω)qnji (ω) − xij (, ω)qnij (ω)⎦ d() + D̄i (t, ω) + zi (t, ω),
[0,t)
j∈B j∈B j∈B j∈B

 − a.e.,∀t ∈ [0, T ], ∀i ∈ B

and:
⎡ ⎤
    
xij (t, ω)qnij (ω) ≤ B0i + ⎣ xji (, ω)qnji (ω) − xij (, ω)qnij (ω)⎦ d() + D̄i (t, ω) + zi (t, ω) + xij (t, ω)pnij (ω)
[0,t)
j∈B j∈B j∈B j∈B

Since pn → 0 strongly, we have that, ∀ε > 0, ∃n0 such that, ∀n ≥ n0 , pn L1 ≤ (ε/(T (1 + sup x))). Then:
⎡ ⎤
   
xij (t, ω)qnij (ω) − ⎣ xji (, ω)qnji (ω) − xij (, ω)qnij (ω)⎦ d() ≤ B0i + D̄i (t, ω) + zi (t, ω) + ε
[0,t)
j∈B j∈B j∈B

The left-hand side converges weakly to:

⎡ ⎤
   
xij (t, ω)qij (ω) − ⎣ xji (, ω)qji (ω) − xij (, ω)qij (ω)⎦ d()
[0,t)
j∈B j∈B j∈B

so that:
⎡ ⎤
    ε
xij (t, ω)qij (ω) ≤ B0i + ⎣ xji (, ω)qji (ω) − xij (, ω)qij (ω)⎦ d() + D̄i (t, ω) + zi (t, ω) +
[0,t)
2
j∈B j∈B j∈B

Since ε is arbitrary, we have that q ∈

(t, q), i.e., 0 ∈ ˚t (q). Therefore, ˚t is demiclosed at 0. 

Lemma 8.
: [0, T ] × K  K is nonempty, convex, and closed-valued.

Proof of Lemma 2. By definition, the set-valued map

: [0, T ] × K  K is given by
⎧ ⎡ ⎤
⎨    

(t, q) = ∈K : xij (t, ω) ij (ω) ≤ B0i + ⎣ xji (, ω)qji (ω) − xij (, ω)qij (ω)⎦ d() + D̄i (t, ω) + zi (t, ω),
⎩ [0,t)
j∈B j∈B j∈B
⎫
⎬
 −a.e.,∀i ∈ B
⎭

Since 0 ∈
(t, p), it is nonempty. Convexity and closedness of
(t, q) follow from the fact that it is defined in terms of linear
inequalities. 

Lemma 9. Let FIX(

) be the set of random fixed points of
: [0, T ] × K  K. Then FIX(
) is convex.
 R. Peñaloza / Journal of Mathematical Economics 45 (2009) 679–692 691

Proof of lemma 3. Let ,  ∈ FIX(

). Then (t) ∈
(t, (t)) and  (t) ∈
(t,  (t)). Let 0 < ˛ < 1 and define  = ˛ + (1 −
˛)  . Then:
 

xij (t, ω) ij (ω) = xij (t, ω)(˛ ij (ω) + (1 − ˛) ij (ω))
j∈B j∈B
 
=˛ xij (t, ω) ij (ω) + (1 − ˛) xij (t, ω) ij (ω) ≤ ˛{B0i
j∈B j∈B
⎡ ⎤
  
+ ⎣ ji (t, ω)xji (, ω) − ij (t, ω)xij (, ω)⎦ d() + D̄i (t, ω) + zi (t, ω)}
[0,t)
j∈B j∈B
⎡ ⎤
  
+(1 − ˛){B0i + ⎣ ji (t, ω)xji (, ω) − ij (t, ω)xij (, ω)⎦ d() + D̄i (t, ω) + zi (t, ω)}
[0,t)
j∈B j∈B
⎡ ⎤
    
= B0i + ⎣ ji (t, ω)xji (, ω) − ij (t, ω)xij (, ω)⎦ d() + D̄i (t, ω) + zi (t, ω)
[0,t)
j∈B j∈B

Then  ∈ FIX(
), so FIX(
) is convex. 

We will now prove the main theorem.

Theorem. Assume (A1) and (A2). Then there exists an optimal settlement function, ∗ , that solves (P), i.e., arg sup{x · :
/ ∅. Moreover, ∗ satisfies x ∈ NFIX(
) ( ∗ ), where NFIX(
) ( ) is the normal cone to FIX(
) at ∗ .
∈ FIX(
)} =

Proof of the theorem. First we will show that FIX(

) = / ∅. By Lemma 1, Id(·)-
(t, ·) is demiclosed at 0, ∀t ∈ [0, T ]. Together
with (A2), this implies that FIX(
) is measurable (Xu and Beg, 1998), hence
has a random fixed point, i.e., FIX(
) = / ∅.
Second, we will show that FIX(
) is closed. Let { m }m≥1 ⊂ FIX(
) be a sequence of random fixed points such that m → .
Then m (t) ∈
(t, m (t)), ∀t ∈ [0, T ], ∀m ≥ 1. Define ˚ : K  K by ˚(x) = Id(x) −
(t, x). Consider ˚( m (t)) = Id( m (t)) −

(t, m (t)). Fix m ∈ FIX(

) and define the set-valued map Am : [0, T ]  K by Am (t) =
(t, m (t)). By Lemma 2, Am (t) is
closed, hence measurable. Since Am is measurable, it follows from the Castaing’s representation theorem that there exists
a sequence {fm,k }k≥1 of measurable selectors of Am such that cl∪k≥1 {fm,k (t)} = Am (t), ∀t ∈ [0, T ], where cl denotes the closure
of a set. Since fm,k (t) ∈ Am (t), ∀t ∈ [0, T ], ∀k ≥ 1, and since m (t) ∈ Am (t), we can extract a subsequence fm,kj such that fm,kj →
m . Define ym,j = m − fm,kj and pick the diagonal subsequence ym = ym,m = m − fm,km . By construction, ym ∈ Id( m (t)) −

(t, m (t)) and ym → 0. Since, by Lemma 1,

(t, ·) is demiclosed at 0, ∀t ∈ [0, T ], we have that 0 ∈ Id( (t)) −
(t, (t)),
∀t ∈ [0, T ], wich implies that (t) ∈
(t, (t)), that is, ∈ FIX(
). Therefore,FIX(
) is closed.
Consider W = L1 (, K) ⊂ L1 (, R ). Obviously, W is bounded and lim(S)→0 S f  d = 0 uniformly in f ∈ W . Fix S ∈ B([0, T ]).

Clearly the set { S f d : f ∈ W } is relatively weakly compact. Then by Dunford’s theorem (Diestel and Uhl, 1977, p. 101), W is
relatively weakly compact. Finally, since FIX(
) ⊂ W is closed, we have that FIX(
) is relatively weakly compact as well.
Then x · achieves a maximum on FIX(
), i.e., arg sup{x · : ∈ FIX(
)} = / ∅.
Consider again the Central Bank’s liquidity problem, sup ∈ L1 (,K) {x · : ∈ FIX(
)}. The indicator function of FIX(
) ⊂
W is given by

0, if ∈ FIX(
)
FIX(
) ( ) =
∞, if ∈
/ FIX(
)

Then (P) is equivalent to

inf {−x · + FIX(

) ( )}
∈ L1 (,K)

Let f ( ) = −x · + FIX(
) ( ). A settlement function ∗ is optimal if 0 ∈ ∂f ( ∗ ) = −x + NFIX(
) ( ∗ ). In other words, ∗ is an
optimal settlement function if x ∈ NFIX(
) ( ∗ ). Here:

NFIX(
) ( ) = {ς ∈ L∞ (, E ∗ ) : ς · ≤ 0, ∀ ∈ TFIX(
) ( ∗ )}

is the normal cone to FIX(

) at ∗ and TFIX(
) ( ∗ ) = { ∈ L1 (, E) : dFIX(

o ( ∗ ; ) = 0} is the tangent cone to FIX(

) at ∗ ,
)
where
dFIX(
) (y + ı ) − dFIX(
) (y)
o
dFIX(
)
( ∗ ; ) = lim sup

y→ ∗ ,ı↓0 ı
692 R. Peñaloza / Journal of Mathematical Economics 45 (2009) 679–692

and dFIX(
) (y) = inf {y −  :  ∈ FIX(
)} (see Hu and Papageorgiou, 1997; Papageorgiou, 1982). 

Proposition. The operators ˚ and  satisfy , ˚ = , , hence  = ˚∗ , i.e.,  is the adjoint of ˚.

Proof. Applying Fubini’s theorem, we have:

 T
, ˚ = (t, ω)X(t, ω) (t, ω) − (t, ω)˚( )(t, ω) d(t)
0
 T  t 
= (t, ω)X(t, ω) (t, ω) − (t, ω) [Y (, ω) − X(, ω)] (, ω) d() d(t)
0 0
 T  t
= (t, ω)X(t, ω) (t, ω) − (t, ω)[Y (, ω) − X(, ω)] (, ω) d() d(t)
0 0

Define K(t, , ω) = (t, ω)[Y (, ω) − X(, ω)] (, ω). Notice that, for each fixed ω ∈ ˝:
 T  t  T  T
K(t, , ω) d() d(t) = K(, t, ω) d() d(t)
0 0 0 t

It is easy to see that K(·, ·, ω) ∈ L1 ([0, T ] × [0, T )). Indeed, the components
 of X and Y are measurable and bounded,
and is bounded by 1. The integral on the left-hand side is equal to K(t, , ω) d() d(t), where  ⊂ [0, T ] × [0, T ]
 
is the triangle  = {(t, ) ∈ [0, T ] × [0, T ] : 0 ≤  ≤ t, 0 ≤ t ≤ T }. Simple calculations9 show that K(t, , ω) d() d(t) =
TT 

0 t
K(, t, ω) d() d(t). Thus:
 T  t
, ˚ = (t, ω)X(t, ω) (t, ω) − (t, ω)[Y (, ω) − X(, ω)] (, ω) d() d(t)
0 0
 T  T
= (t, ω)X(t, ω) (t, ω) − (t, ω)[Y (, ω) − X(, ω)] d() (, ω) d(t) = ,
0 t

Therefore,  = ˚∗ , i.e.,  is the adjoint operator of ˚, as was to be shown. 

References

Anderson, E., Nash, P., 1987. Linear Programming in Infinite-Dimensional Spaces. John Wiley & Sons, New York.
Angelini, P., 1998. An analysis of competitive externalities in gross settlement systems. Journal of Banking and Finance 22, 1–18.
Bank for International Settlements, 1997. Real-time gross settlement systems. CPSS, March, Basle, Switzerland.
Bech, M., Garrat, R., 2001. The intraday liquidity management. Department of Economics, University of California, Santa Barbara (UCSB), mimeo.
Bech, M., Soramäki, K., 2001. Gridlock resolution in interbank payment systems. Bank of Finland Discussion Paper 9-2001.
DeBandt, O., Hartmann, P., 2000. Systemic risk: a survey. European Central Bank, WP # 35.
Diestel, J., Uhl, J., 1977. Vector Measures. American Mathematical Society, Mathematical Surveys and Monographs # 15, Providence, RI.
Freixas, X., Parigi, B., 1998. Contagion and efficiency in gross and net interbank payment systems. Journal of Financial Intermediation 7, 3–31.
Gretsky, N., Zame, W., Ostroy, J., 1992. The nonatomic assignment model. Economic Theory 2, 103–127.
Gretsky, N., Zame, W., Ostroy, J., 1999. Perfect competition in the continuous assignment model. Journal of Economic Theory 88, 60–118.
Gretsky, N., Zame, W., Ostroy, J., 2002. Subdifferentiability and the duality gap. Positivity 6, 261–274.
Güntzer, M., Jungnickel, D., Leclerc, M., 1998. Efficient algorithms for the clearing of interbank payments. European Journal of Operational Research 106,
212–219.
Hu, S., Papageorgiou, N., 1997. Handbook of Multivalued Analysis. I. Theory. Kluwer Academic Publishers, Dordrecht, The Netherlands.
Leinonen, H., Soramäki, K., 1999. Optimizing liquidity usage and settlement speed in payment systems. Bank of Finland discussion paper # 16/99.
Makowski, L., Ostroy, J., 2000. Linear programming and general equilibrium theory. In: First World Congress of the Game Theory Society (Games 2000),
Bilbao, Spain, July 24–28.
McAndrews, J., Rajan, S., 2000. The timing and funding of FEDWIRE funds transfers. FRBNY Economic Policy Review, July. Federal Reserve Bank of New York.
Papageorgiou, N., 1982. A class of infinite dimensional linear programming problems. Journal of Mathematical Analysis and Applications 87, 228–245.
Peñaloza, R., 2002. Shadow-prices in payment systems. Working Paper #241, Department of Economics, University of Brasilia.
Peñaloza, R., 2003. On shadow-prices of banks in real-time gross settlement systems. Working Paper Series, 71, Banco Central do Brasil.
Peñaloza, R., 2005. Preços-sombra no sistema de pagamentos: uma abordagem dual para a política monetária intradiária. Revista Brasileira de Economia
59 (4), 609–634.
Roberds, W., 1999. The incentive effects of settlement systems: a comparison of gross settlement, net settlement, and gross settlement with queueing. IMES
discussion paper #99-E-25, Bank of Japan.
Schechter, M., 1972. Linear programs in topological vector spaces. Journal of Mathematical Analysis and Applications 37, 492–500.
Tyndall, W., 1967. An extended duality theorem for continuous linear programming problems. SIAM Journal of Applied Mathematics 15, 1294–
1298.
Xu, H-K., Beg, I., 1998. Measurability of fixed point sets of multivalued random operators. Journal of Mathematical Analysis and Applications 225, 62–72.

9
Solve this integral as an iterated integral in the reverse order and interchange the names of t and  (see Schechter, 1972, lemma 2.2).