Fragmentation and optimal liquidity supply on

decentralized exchanges

Alfred Lehar
Christine A. Parlour
Marius Zoican∗

May 12, 2024

Abstract

We investigate how liquidity providers (LPs) choose between trading venues with high and
low fees, in the face of a fixed common gas cost. Analyzing Uniswap data, we find that high-fee
pools attract 58% of liquidity supply but execute only 21% of trading volume. Large LPs
dominate low-fee pools, frequently adjusting positions in response to substantial trading volume.
In contrast, small LPs converge to high-fee pools, accepting lower execution probabilities to
mitigate smaller liquidity management costs. Fragmented liquidity dominates a single-fee market,
as it encourages more liquidity providers to enter the market, while enhancing LP competition
on the low-fee pool.

Keywords: FinTech, decentralized exchanges (DEX), liquidity, fragmentation, retail trading

JEL Codes: G11, G12, G14

∗
Alfred Lehar ([Link]@[Link]) is affiliated with Haskayne School of Business at University
of Calgary. Christine A. Parlour (parlour@[Link]) is with Haas School of Business at UC Berkeley. Marius
Zoican ([Link]@[Link], corresponding author) is affiliated with University of Toronto Mississauga
and Rotman School of Management. Corresponding address: 3359 Mississauga Road, Mississauga, Ontario L5L 1C6,
Canada. We have greatly benefited from discussion on this research with Michael Brolley, Agostino Capponi (discussant),
Itay Goldstein, Sang Rae Kim (discussant), Olga Klein, Katya Malinova (discussant), Ciamac Moallemi, Uday Rajan,
Thomas Rivera, Andreas Park, Gideon Saar, Lorenzo Schöenleber (discussant), Andriy Shkilko (discussant), and Shihao
Yu. We are grateful to conference participants at the 2024 NYU Stern Market Microstructure Meeting, Tokenomics
2023, Gillmore Centre Annual Conference 2023, Edinburgh Economics of Technology, Financial Intermediation Research
Society 2023, the Northern Finance Association 2023, the UNC Junior Faculty Finance Conference, as well as to
seminar participants at the University of Chicago, Lehigh, the Microstructure Exchange, UCSB-ECON DeFi Seminar,
University of Melbourne, Wilfrid Laurier University, University of Guelph, Rotman School of Management, Hong
Kong Baptist University, and Bank of Canada. Marius Zoican gratefully acknowledges funding support from the
Rotman School of Managements’ FinHub Lab and the Canadian Social Sciences and Humanities Research Council
(SSHRC) through an Insight Development research grant (430-2020-00014).
 Fragmentation and optimal liquidity supply on
decentralized exchanges

Abstract

We investigate how liquidity providers (LPs) choose between trading venues with high and
low fees, in the face of a fixed common gas cost. Analyzing Uniswap data, we find that high-fee
pools attract 58% of liquidity supply but execute only 21% of trading volume. Large LPs
dominate low-fee pools, frequently adjusting positions in response to substantial trading volume.
In contrast, small LPs converge to high-fee pools, accepting lower execution probabilities to
mitigate smaller liquidity management costs. Fragmented liquidity dominates a single-fee market,
as it encourages more liquidity providers to enter the market, while enhancing LP competition
on the low-fee pool.

Keywords: FinTech, decentralized exchanges (DEX), liquidity, fragmentation, retail trading

JEL Codes: G11, G12, G14
1 Introduction

In addition to aggregating information, asset markets allow agents to exhaust private gains from
trade. While there is a well developed literature on the informativeness of prices, less is known about
if and when markets effectively exhaust all gains from trade.1 In this paper, we exploit the unique
design of a decentralized exchange to shed light on the market for liquidity and show, theoretically
and empirically, that market fragmentation can improve trading efficiency.
Automated market makers such as Uniswap v3 provide a unique environment to investigate the
market for liquidity. While there are various new institutional details that animate these exchanges,
for our purposes three are economically important. First, in automated exchanges, liquidity demand
and supply can easily be distinguished: users either supply or demand liquidity. Because of this,
we can isolate the effect of transactions costs on each side of the market for liquidity. Second,
costs and benefits incurred by liquidity suppliers are easier to observe because prices are not set by
market participants, but are automatically calculated as a function of liquidity demand and supply.
Thus, liquidity suppliers are not compensated through price impact. Third, market participants are
pseudo-anonymous so we can identify and document liquidity suppliers at a high frequency. These
unique features allow us to investigate, theoretically and empirically, how transactions costs affect
liquidity supply.
Beyond investigating the market for liquidity, there are three additional reasons to investigate
liquidity provision in AMMs. First, these markets are large and successful in their own right: After
its May 2021 launch, Uniswap v3 features daily trading volume in excess of US $1 billion. Second,
for major pairs such as Ether against USD stablecoins, Uniswap boasts twice or three times better
liquidity than continuous limit order exchanges such as Binance, which suggests that this design
can be economically superior.2 Third, as traditional assets become tokenized, and markets become
more automated, this new market form could be adopted.3
Uniswap v3 provides two innovations over the previous v2. First, liquidity suppliers and
demanders select into trading places (called pools) that differ on transaction fees. Each asset pair to
be traded on up to four liquidity pools that only differ in the compensation for liquidity providers:
in particular, liquidity fees can be equal to 1, 5, 30, or 100 basis points and the corresponding tick
sizes are 1, 10, 60, or 200 basis points. These proportional fees are paid by liquidity demanders
and are the only source of remuneration to liquidity providers. (These fees, as we discuss below,
are similar to the make-take fees that are prevalent in limit order markets.) Second, on Uniswap
1
Gains from trade comprise an idiosyncratic private value for the underlying asset, but also idiosyncratic preferences
for trade speed or “liquidity.”Agents’ idiosyncratic value for the underlying asset are plausibly determined by their
portfolio positions, and therefore independent of the trading place. By contrast, their idiosyncratic preference for
liquidity determines market quality.
2
See The Dominance of Uniswap v3 Liquidity; May 5, 2022.
3
Swarm — a BaFin regulated entity — already offers AMM trading for a variety of tokenized Real World Assets.

1
v3, liquidity providers can submit “concentrated liquidity.” Even though their liquidity is passively
supplied, they can choose the price range over which it is supplied. With volatile assets, these
concentrated liquidity positions can become stale and require rebalancing.
Besides differences in fees, the liquidity pools are otherwise identical and, importantly, they
share the common infrastructure of the Ethereum blockchain. Importantly, all participants pay a
transaction cost (called a “gas fee”) to access the markets. Our theory and empirical work investigates
the effect of different proportional fees and fixed fees on liquidity supply. At launch, Uniswap Labs
conjectured that trading and liquidity should consolidate in equilibrium on a single “canonical” pool
for which the liquidity fee is just enough to compensate the marginal market maker for adverse
selection and inventory costs. That is, activity in low-volatility pairs such as stablecoin-to-stablecoin
trades should naturally gravitate to low fee liquidity pools, whereas speculative trading in more
volatile pairs will consolidate on high fee markets.4 As we show, this reasoning is flawed.
We present a simple model with trade between liquidity suppliers and two types of liquidity
demand. Consistent with the design of v3, liquidity suppliers chose a market and then place their
liquidity into a band around the current value of the asset. The posted liquidity is subject to a
bonding curve and hence generates a price impact cost for the liquidity demanders (we emphasize
that this does not remunerate the liquidity suppliers). Liquidity suppliers have heterogeneous
endowments, interpretable as different capital constraints — low-endowment liquidity providers
are akin to retail traders, whereas high-endowments stand in for large institutional investors or
quantitative funds. Trade occurs against these positions either because a liquidity demander
arrives who has experienced a liquidity shock or because the value of the asset has changed and an
arbitrageur adversely selects the passive liquidity supply. Collectively, the decisions of the liquidity
demanders determine the payoff to the liquidity suppliers. After large private or common value
trades, liquidity providers rebalance their positions; to do so, liquidity providers incur a fixed cost
(i.e., gas price) each time they update their position.
Traders demanding liquidity face two types of costs: first, the fee associated with their chosen
pool (low or high) and second, the price impact costs generated by the pool’s bonding curve and
supplied liquidity. They route their orders to the all-in lowest cost trading venue. We find that for
small orders, the low-fee pool minimizes transaction costs. The high-fee pool executes the residual
order flow from large traders who exhaust cheap liquidity at the low-fee venue. As a result, low-fee
markets are actively traded and require frequent liquidity updates whereas high-fee pools have a
longer liquidity update cycle since they absorb fewer trades.
We establish conditions under which there is fragmentation or consolidation. Specifically, even
in this simple framework, there is a robust parameter range in which liquidity does not naturally
concentrate on one of the exchanges. Both pools can attract a positive market share if the gas
4
See Flexible fees paragraph at [Link] accessed September 14, 2022.

2
price is large enough due to economies of scale. Liquidity providers trade off a higher revenue per
unit of time in the low-fee pool (driven by the larger trading volume) against the additional gas
cost required for active liquidity management. As a result, liquidity provider clienteles emerge in
equilibrium. Liquidity providers with large endowments gravitate towards low-fee markets, as they
are best positioned to frequently update their position. In contrast, smaller market makers choose
to passively provide liquidity on high-fee markets where they only trade against large orders being
routed there. They optimally trade off a lower execution probability against higher fees per unit of
volume and a lower liquidity management cost per unit of time.
Not only does liquidity fragment, but it differs in both use and type across the two markets. A
small number of highly active large liquidity providers, potentially institutional investors and hedge
funds, primarily trade against numerous small incoming trades on pools with low fees. In contrast,
high-fee pools involve less frequent trading between a substantial number of capital-constrained
passive liquidity providers (e.g., retail market makers) on one side and a few sizeable incoming orders
on the other. As the fixed gas fee affects liquidity providers pool choice, changes in the common
fixed market access fee differentially affects the liquidity supply on the two pools. Specifically, it
reduces market quality (in the sense of lower posted liquidity) on the low fee pool.
As we distinguish between liquidity demanders who are trading to exploit gains from trade
and liquidity demanders who are arbitraging common value changes, we can decompose returns to
liquidity providers, and show that adverse selection is higher on the low fee pool. Given that the
low fee pool is populated with larger liquidity suppliers, this suggests that institutional traders bear
price risk.
Using the model for guidance, we analyze more than 28 million interactions with Uniswap v3
liquidity pools – that is, all liquidity updates and trades from the inception of v3 in May 2021 until
July 2023. We first document liquidity fragmentation in 32 out of 242 asset pairs in our sample,
which account for 95% of liquidity committed to Uniswap v3 smart contracts and 93% of trading
volume. For each of the fragmented pairs, trading consolidates on two pools with adjacent fee levels:
either 1 and 5 basis points (e.g., USDC-USDT), 5 and 30 basis points (ETH-USDC), or 30 and 100
basis points (USDC-CRV).
We then document that high-fee pools are on average larger – with aggregate end-of-day liquidity
of $46.50 million relative to $33.78 million, the average size of low-fee pools. Nevertheless, three
quarters of daily trading volume executes on low-fee pools. In line with the model predictions,
low-fee pools are more active as they capture many small trades. There are five times as many
trades on low- than on high-fee pools (610 versus 110). However, the average trade on the high fee
pool is twice as large: $14,490 relative to $6,340. Unsurprisingly, liquidity cycles – measured as the
time between the submission and update of posted liquidity – are 20% shorter on the highly active
low-fee pool.

3
 We find robust evidence of liquidity supply clienteles across pools. The average liquidity deposit
is 107.5% larger on the low-fee pool, after controlling for daily volume and return volatility. At the
same time, high-fee pools’ market share is 21 percentage points higher. The results point to an
asymmetric match between liquidity supply and demand: large liquidity providers are matched with
small liquidity demanders on low-fee pools, whereas small liquidity providers trade with a few large
orders on the high-fee pool.
We then turn to the common fixed cost of accessing the market, or gas fees. The market shares
of the liquidity pools depend on the magnitude of gas costs on the Ethereum blockchain. In the
model, a higher gas price leads to a shift in liquidity supply from the low- to the high-fee pool as
active position management becomes relatively more costly for the marginal liquidity provider. We
find that a one standard deviation increase in gas prices corresponds to a 4.63 percentage points
decrease in the low-fee pool market share, and a 29% drop in liquidity inflows on days when gas
costs are elevated. With fewer liquidity providers on the low-fee pool, the time between position
updates becomes 4.19% shorter since incoming order flow depletes liquidity at a faster pace. The
effect reinforces the expected gas cost difference between the two pools, and more liquidity providers
switch to the high-fee venue.
Consistent with our model, the impact of gas prices on liquidity supply is not driven by variation
in price impact, as measured by loss-versus-rebalancing (LVR, as in Milinois, Moallemi, Roughgarden,
and Zhang, 2023) and impermanent loss (Heimbach, Schertenleib, and Wattenhofer, 2022). While
low-fee pools face a greater permanent price impact, differences between high- and low-fee pool
deviations from hourly centralized exchange prices are not significant.
Our paper is related to various literatures. Pagano (1989) shows that if an asset is traded on
two identical exchanges with equal transaction costs, in equilibrium market participants gravitate to
a single exchange due to network effects. In practice, exchanges are rarely identical: fragmentation
can emerge between fast and slow exchanges (Pagnotta and Philippon, 2018; Brolley and Cimon,
2020) or between lit and dark markets (Zhu, 2014). In our model, fragmentation on decentralized
exchanges is driven by variation in liquidity fees as well as different economies of scale due to
heterogeneity in liquidity provider capital. We find that liquidity fragmentation driven by high gas
fees implies larger transaction costs on incoming orders. We note that there is no time priority on
decentralized exchanges, which clear in a pro rata fashion. On markets with time priority, Foucault
and Menkveld (2008) and O’Hara and Ye (2011) find that market segmentation in equity markets
improves liquidity (by allowing queue jumping) and price discovery.
Fixed costs for order submission are uncommon in traditional markets. However, in 2012, the
Canadian regulator IIROC implemented an “integrated fee model” that charged traders for all
messages sent to Canadian marketplaces. Korajczyk and Murphy (2018) document that this measure
disproportionately affected high-frequency traders, resulting in wider bid-ask spreads but lower
implementation shortfall for large traders, possibly due to a reduction in back-running activity. Our

4
study contributes additional insights by highlighting that the introduction of a fixed cost, even when
applied across exchanges, can lead to market fragmentation.
We also relate to a rich literature on market fragmentation and differential fees. Closest to
our paper, Battalio, Corwin, and Jennings (2016) and Cimon (2021) study the trade-off between
order execution risk and compensation for liquidity provision in the context of make-take fee
exchanges. However, Battalio, Corwin, and Jennings (2016) specifically addresses the issue of the
broker-customer agency problem, whereas our study focuses on liquidity providers who trade on their
own behalf. In traditional securities markets, make-take fees are contingent on trade execution and
proportional to the size of the order. On the other hand, gas costs on decentralized exchanges are
independent of order execution, highlighting the significance of economies of scale (lower proportional
costs for larger liquidity provision orders) and dynamic liquidity cycles (managing the frequency
of fixed cost payments). Strategic brokers in Cimon (2021) provide liquidity alongside exogenous
market-makers in a static setting. We complement this approach by modelling network externalities
inherent in the coordination problem of heterogeneous liquidity providers. In our dynamic setup,
this allows us to pin down the equilibrium duration of liquidity cycles and the relative importance
of gas fixed costs.
Our paper relates to a nascent and fast-growing literature on the economics of decentralized
exchanges. Many studies (e.g., Aoyagi, 2020; Aoyagi and Ito, 2021; Park, 2022) focus on the
economics of constant-function automated market makers, which do not allow liquidity providers to
set price limits. In this restrictive environment, Capponi and Jia (2021) argue that market makers
have little incentives to update their position upon the arrival of news to avoid adverse selection,
since pro-rata clearing gives an advantage to arbitrageurs. Lehar and Parlour (forthcoming) solve
for the equilibrium pool size in a setting where liquidity providers fully internalize information
costs without rushing to withdraw positions at risk of being sniped. We argue that on exchanges
that allow for limit or range orders, the cost of actively managing positions becomes a first-order
concern, as liquidity providers need to re-set the price limits once posted liquidity no longer earns
fees. Our empirical result on economies of scale echoes the argument in Barbon and Ranaldo (2021),
who compare transaction costs on centralized and decentralized exchanges and find that high gas
prices imply that the latter only become competitive for transactions over US$100,000. Hasbrouck,
Rivera, and Saleh (2022) argue that liquidity providers require remuneration. We complement the
argument by stating that high fees might be necessary for some liquidity providers to cover the fixed
costs of managing their position. In line with our theoretical predictions, Caparros, Chaudhary,
and Klein (2023) find that liquidity providers reposition their quotes more often on Uniswap V3
pools built on Polygon, which features substantially lower gas fees. Finally, Heimbach, Schertenleib,
and Wattenhofer (2022) document that after accounting for price impact, concentrated liquidity
on Uniswap v3 pools results in increased returns for sophisticated participants but losses for retail
traders.

5
 Despite higher gas costs, decentralized exchanges may hold advantages over centralized venues.
Han, Huang, and Zhong (2022) demonstrate Uniswap frequently leads price discovery compared to
centralized exchanges such as Binance, despite the latter having higher trading volume. Capponi,
Jia, and Yu (2023) find that the fee paid by traders to establish execution priority unveils their
private information, and therefore contributes to price discovery. Aspris, Foley, Svec, and Wang
(2021) argue that decentralized exchanges offer better security than their centralized counterparts
since assets are never transferred to the custody of a third party such as an exchange wallet. In
turn, Brolley and Zoican (2023) make the point that decentralized exchanges may be able to reduce
overall computational costs associated with latency arbitrage races, as they eliminate long-term
co-location subscriptions.
Our paper is related to both the finance literature that examines whether make-take fees affect
market quality and to the economics literature on two-sided markets and platform competition.
Broadly, our work differs from the finance literature in that we explicitly consider equity markets as
markets for liquidity without focusing on the order choice decision, and our work differs from the
economics literature in that we explicitly analyze an equity market as a market for liquidity. The
main insight that this brings is that market participants are both large and strategic, compared to
smaller players in consumer-facing markets that are often analyzed in the economics literature.

2 Model

Asset and agents. Consider a continuous time model of trade in a single token T with expected
value vt > 0. Three risk neutral trader types consummate trade in this market: a continuum of
liquidity providers (LPs), liquidity traders (LTs), and arbitrageurs (A). Trade occurs either because
public news triggers a change in the common value of the asset, or because market participants
have heterogeneous private values for the asset.
Arrival times of news and private value shocks follow independent Poisson processes with rates
η ∈ (0, 1) and 1 − η, respectively.5 For notational compactness, we first characterize the generic
shock distribution and then describe its effects onarbitrageurs
 or liquidity traders. Conditional on
an event at time t, the asset value changes to vt 1 + I δ̃ for all traders in the case of a common
value shock, or for an arriving (LT) in the case of the private value shock. Here, I is an indicator
that takes on the value of 1 if the taker buys and −1 if the taker sells. The value innovation δ̃ has a
probability density
1
for δ ∈ 0, ∆2 − 1 ,
 
ϕ (δ) = √ (1)
2∆ 1 + δ
5
This is without loss of generality, as what matters in the model is the relative arrival rate of news relative to
liquidity traders.

6
 p
thus 1 + δ̃ is uniformly distributed between [1, ∆]. This assumption is innocuous and made for
tractability purposes.
When news arrives, the innovation is to the common value of the token. (As we are agnostic as
to the source of value of cryptocurrencies, this common value shock could include the possibility
of resale on another exchange.) After such a shock, an arbitrageur A trades with the liquidity
providers whenever profitable, and LPs face an adverse selection loss. Conversely, when a liquidity
trader enters the market, they experience a private value shock — and liquidity providers continue
to value the token at vt . In what follows for expositional simplicity, as in Foucault, Kadan, and
Kandel (2013), we focus on a one sided market in which liquidity takers act as buyers, and news
lead to an increase in token value.
Liquidity providers (LP) differ in their endowments of the token. Each provider i can supply
at most qi di of the token, where qi follows an exponential distribution with scale parameter λ.
The right skew of the distribution captures the idea that there are many low-endowment liquidity
providers such as retail traders, but few high-capital LPs such as sophisticated quantitative funds.
Heterogeneity in LP size is captured by λ, where a larger λ naturally corresponds to a larger
dispersion of endowments and larger aggregate liquidity supply. Given the endowment distribution,
collectively LPs supply at most Z ∞
1 qi
S= qi e− λ di = λ (2)
0 λ
tokens.

Trading environment. Traders can interact in two liquidity pools in which token trade occurs
against a numéraire asset (cash). At the start of the trading game, each liquidity provider (LP)
deposits
h liquidityi to a single pool within a symmetric price band around the current asset value
v
(1+r)2
, v(1 + r)2 , where r ≥ 0. Here, we make use of the fact that V3 features “price bands,” and
thus liquidity can be consumed with a bounded price impact. Within this range, prices in both pools
satisfy a constant product bonding curve as in Adams, Zinsmeister, Salem, Keefer, and Robinson
(2021). In particular, for pool k,
   √ 
Lk v
Tk + √ Tk v + Lk = L2k , (3)
v (1 + r) 1+r
| {z } | {z }
virtual token reserves virtual numeraire reserves

where Tk is the amount of tokens deposited on pool k and Lk is the liquidity level of pool k, defined
as
Tk
Lk = . (4)
√1 − √1 (1 + r)
v v

7
 v(1+r)
To purchase τ tokens, a trader needs to deposit an amount n (τ ) = τ Tk τ +(1+r)(Tk −τ )
of numéraire
into the pool, where n (τ ) is the solution to the invariance condition
  √ 
Lk v
Tk − τ + √ Tk v + n (τ ) + Lk = L2k . (5)
v (1 + r) 1+r
| {z }| {z }
virtual token reserves virtual numeraire reserves

Fees are levied on liquidity takers as a fraction of the value of the trade and distributed pro
rata to liquidity providers. Crucially, the pools have different fees. One pool charges a low fee, and
one pool charges a high fee which we denote ℓ and h respectively. Specifically, to purchase τ units
of the token on the low fee pool, the total cost to a taker is (1 + ℓ) n (τ, Tℓ ). The LPs in the pool
receive ℓn (τ, Tℓ ) in fees. In addition, consistent with gas costs on Ethereum, all traders incur a
fixed execution cost Γ di > 0 to interact with the market.
Figure 1 illustrates the timing of the model.

A trades LP rebalance
Small Small Large A trades LP rebalance Large Small on both on both
LT LT LT on pool L on pool L LT LT pools pools Pool L

LPs deposit qi News News

in pool k ∈ {L, H} δ̃ ∈ (ℓ, h) δ̃ > h Pool H

Figure 1: Model timing

To ensure the possibility of trade in both pools, we assume that innovations are large enough to
ensure that trade could be feasible on the high fee pool or:

Assumption 1: The size of innovations are sufficiently large so that there is a positive probability
√
that liquidity providers need to rebalance on the high-fee pool. That is, ∆ > (1 + r) 1 + h.

2.1 Equilibrium

First, consider the decisions of arbitrageurs and liquidity traders holding a value v (1 + δ) for the
asset. Faced with pool sizes of Tℓ and Th in the low and high pool respectively, their optimal trade
on pool k maximizes their expected profit, net of fees and price impact:

v (1 + r)
max Profit LT (τ, δ) ≡ τ v (1 + δ) − (1 + fk ) τ Tk , (6)
τ τ + (1 + r) (Tk − τ )

8
which yields the optimal trade quantity:
( ( r ))
1 + r 1 + fk
τ ⋆ (δ) = Tk min 1, max 0, 1 − . (7)
r 1+δ

From equation (7), a trader with valuation v (1 + δ) only trades on pool k if the gains from trade are
larger than the liquidity fee, i.e., δ > fk . Further, if δ > (1 + fk ) (1 + r)2 − 1 so that the gains from
trade are larger than the maximum price impact, then the trader consumes all available liquidity in
the pool.
Therefore, the aggregate fee revenue for liquidity providers is equal to
( ( s ))
1 + r 1 + δ
fk n (τ ⋆ , Tk ) = fk vTk min 1 + r, max 0, −1 . (8)
r 1 + fk

If the innovation is a private value shock (i.e., coming from an LT), an arbitrageur optimally
steps in to reverse the liquidity trade (as in Lehar and Parlour, forthcoming). In this case, the
liquidity providers earn twice the fee revenue and since the pool is back to the original state there is
no capital gain or loss.
We note that the total revenue scales linearly with the size of the pool, and therefore the fee
revenue for LP with endowment qi providing liquidity on pool k is simply:

qi
ProfitLiqi,k = 2 fk n (τ ⋆ , Tk )
Tk
( ( s ))
1+r 1+δ
= 2qi vfk min 1 + r, max 0, −1 . (9)
r 1 + fk

We take the expectation of fee revenues over the size of private value shocks δ and obtain:
n
EProfitLiqi,k = 2qi vfk ×
"s #
1 + r 1 + δ
P fk < δ ≤ (1 + fk )(1 + r)2 − 1 × − 1 | fk < δ ≤ (1 + fk )(1 + r)2 − 1 +


E
r 1 + fk
o
+ P δ > (1 + fk )(1 + r)2 − 1 × (1 + r)


√ √

fk (r + 1) 2∆ − r fk + 1 − 2 fk + 1
= qi v , (10)
| ∆
{z }
≡L(fk )

where we define L (fk ) as the per-unit profit from liquidity provision in pool k.

Lemma 1. There exists a threshold fee level f > 0 such that the liquidity revenue L (fk ) first increases

9
in the pool fee fk for f ≤ f , then decreases in the pool fee for f > f .

A salient implication of Lemma 1 is that if pool fees are large enough, the liquidity yield on the
high fee pool may exceed the yield on the low-fee pool.
If the innovation is common value and an arbitrageur trades in the direction of news, there is no
price reversal. In other words, liquidity providers earn fee revenue but also make a capital loss by
trading against the common value innovation. Further, if the arbitrage trade is large enough that
it exhausts all available liquidity, LPs pay the gas fee Γ di to re-balance liquidity around the new
current price.
We tabulate below the revenue of LP with endowment qi from selling tokens to arbitrageurs on
pool k, as well as the marked-to-market value of tokens sold. The LP profit for each case is the
difference between the second and third columns, which is always negative since LPs are trading
against the news.

News size Revenue (numeraire) Value of tokens sold

δ ≤ fk 0  0
 i q  q
δ ∈ fk , (1 + fk ) (1 + r)2 − 1 vqi 1+r
r (1 + fk )
1+δ
1+fk −1 vqi 1+r
r (1 + δ) 1 −
1+fk
1+δ

δ > (1 + fk ) (1 + r)2 − 1 vqi (1 + fk ) (1 + r) vqi (1 + δ)

The expected LP profit from trading with arbitrageurs equals −qi v × A (fk ), where A (fk ) is the
per-unit adverse selection cost from liquidity provision in pool k:

1+r h p i
A (fk ) = P fk < δ ≤ (1 + fk )(1 + r)2 − 1 × E (1 + fk ) + (1 + δ) − 2 (1 + δ) (1 + fk ) +
r o
2


+ P δ > (1 + fk )(1 + r) − 1 × [E (1 + δ)] − (1 + fk ) (1 + r) . (11)

Lemma 2. The adverse selection cost A (fk ) decreases in the pool fee fk . In particular, the high-fee
pool has a lower adverse selection cost than the low-fee pool.

10
 Figure 2: Liquidity revenues and adverse selection cost
This figure illustrates the expected fee revenue from liquidity traders (left panel) and the adverse selection√cost (right
panel), as a function of the pool fee f . Parameter values: r = 0.001, λ = 1, η = 0.1, and ∆ = 1.1 (1 + r) 1 + h.

Liquidity providers only need to re-balance if their position is out of range following a news
event. The expected cost of rebalancing following news is:
 √ 
2 1 + fk (1 + r)


C (k) = P δ > (1 + fk )(1 + r) − 1 Γ = Γ 1 − , (12)
∆

which is decreasing in the pool fee fk . Therefore, LPs on lower fee pools re-balance more often and
incur higher fixed costs.
We turn next to the choice of pool for liquidity providers. The expected profit from providing
liquidity on the low- and high-fee pools, respectively, is equal to:
 √ 
1 + ℓ (1 + r)
πL = qi [(1 − η) L (ℓ) − ηA (ℓ)] − ηΓ 1 − and (13)
∆
 √ 
1 + h (1 + r)
πH = qi [(1 − η) L (h) − ηA (h)] − ηΓ 1 − . (14)
∆

We define q L and q H as the unique solutions to πL (q) = 0 and πH (q) = 0, respectively. If q k ≥ 0

for a pool k, then any LP with qi ≥ q earns positive profit by joining pool k, with the marginal
entrant making zero profit. Conversely, if q k < 0, then pool k is not viable since the participation
constraint is violated for all LPs.

Assumption 2: We focus on the case that both markets are potentially viable, or equivalently the
intensity of news is low enough:
L (k)
η ≤ min . (15)
k L (k) + A (k)

11
 Next, we investigate equilibria where the market becomes fragmented. A LP with endowment qi
chooses pool L if:
 
η(1 + r) √ √ 
πL −πH = qi (1 − η) (L (ℓ) − L (h)) +η (A (h) − A (ℓ))−Γ 1+h− 1 + ℓ > 0. (16)
| {z } | {z } ∆
>0 <0

If there exists a fragmented equilibrium, then there exists a threshold qt such that all LPs with
qi > qt choose the low-fee pool and all LPs with qi ≤ qt choose the high-fee pool, where
√ √

η(1 + r 1 + h − 1 + ℓ )
qt = Γ (17)
∆ [(1 − η) (L (ℓ) − L (h)) + η (A (h) − A (ℓ))]

L(l)−L(h)
Proposition 1. i. If η > L(l)−L(h)+A(l)−A(h) , then all LPs with qi > q h deposit liquidity on the
high fee pool.
L(l)−L(h) C(h) (1−η)L(h)−ηA(h)
ii. If η ≤ L(l)−L(h)+A(l)−A(h) and C(ℓ) > (1−η)L(ℓ)−ηA(ℓ) , then all LPs with qi > q ℓ deposit
liquidity on the low fee pool.

iii. Otherwise, there exists a unique fragmented equilibrium characterized by marginal trader
qt⋆ > q h which solves
√ √

η(1 + r 1 + h − 1 + ℓ )
qt⋆ =Γ (18)
∆ [(1 − η) (L (ℓ) − L (h)) + η (A (h) − A (ℓ))]
 i
such that all LPs with qi ∈ q h , qt⋆ deposit liquidity in the high fee pool and all LPs with
qi > qt⋆ choose the low fee pool.

12
 Figure 3: Liquidity supply on fragmented markets
This figure illustrates the endowment distribution of LPs and their choice of pools in a fragmented market. First,
liquidity providers to the left of q h do not provide liquidity on either pool. Next, LPs to the left (right) of the marginal
trader qt⋆ provide
√ liquidity on pool H (pool L, respectively). r = 0.001, h = 2, ℓ = 1, λ = 1, η = 0.1, Γ = 20, and
∆ = 1.1 (1 + r) 1 + h.

Proposition 2 establishes comparative statics for the two pools’ liquidity market shares. We can
compute the liquidity market share of the low-fee pool in a fragmented equilibrium as
 q −q 
t
exp − λ h (qt + λ)
wℓ = ≤ 1, (19)
qh + λ

with equality for Γ = 0.

Proposition 2. In equilibrium, the market share of the low fee pool wℓ decreases in the gas cost (Γ).

Figure 4 shows that the market share of the low fee pool decreases in the gas cost Γ. A larger
gas price increases the costs of re-balancing upon the arrival of large enough news, everything else
equal, and incentivizes smaller LPs to switch from the low fee pool to the high fee pool, since the
arbitrageurs are less likely to fully consume liquidity there. Further, the right panel illustrates the
extensive margin effect of gas prices: any increase in gas costs leads to a decrease in aggregate
liquidity supply as some LP with low endowments are driven out of the market (that is, the threshold
q h increases in Γ).

13
 Figure 4: Liquidity shares and gas costs
This figure illustrates the equilibrium liquidity market shares (left panel) and the aggregate liquidity supply on the
two pools (right√panel), as a function of the gas fee Γ. Parameter values: r = 0.001, h = 2, ℓ = 1, λ = 1, η = 0.1, and
∆ = 1.1 (1 + r) 1 + h.

2.2 Pool fragmentation and market quality

We measure market quality as the realized gains from trade from liquidity traders. across all pools
k, that is k Evτ ⋆ (fk , δ).
P

Th Eτ ′ h, δ ≤ Th Eτ ′ h, δ + Tl Eτ ′ l, δ (20)

3 Data and descriptive statistics

3.1 Sample construction

We obtain data from the Uniswap V3 Subgraph, covering all trades, liquidity deposits (referred
to as “mints”), and liquidity withdrawals (referred to as “burns”) on 4,069 Uniswap v3 pools. The
data spans from the protocol’s launch on May 4, 2021, up until July 15, 2023. Each entry in our
data includes a transaction hash that uniquely identifies each trade and liquidity update on the
Ethereum blockchain. Additionally, it provides details such as trade price, direction, and quantity,
along with quantities and price ranges for each liquidity update. Moreover, the data also includes
wallet addresses associated with initiating each transaction, akin to anonymous trader IDs. The
Subgraph data we obtained also provides USD-denominated values for each trade and liquidity mint.
We further collect daily pool snapshots from the Uniswap V3 Subgraph, including the end-of-day

14
pool size in Ether and US Dollars, and summary price information (e.g., open, high, low, and closing
prices for each pool).
To enhance our dataset, we combine the Subgraph data with public Ethereum data available on
Google Big Query to obtain the position of each transaction in its block, as well as the gas price
limit set by the trader and the amount of gas used.
There are no restrictions to list a token pair on Uniswap. Some pools might therefore be used for
experiments, or they might include untrustworthy tokens. Following Lehar and Parlour (forthcoming),
we remove pools that are either very small or that are not attracting an economically meaningful
trading volume. We retain liquidity pools that fulfill the following four criteria: (i) have at least
one interaction in more than 100 days in the sample, (ii) have more than 500 liquidity interactions
throughout the sample, (iii) have an average daily liquidity balance in excess of US$100,000, and
(iv) capture more than 1% of trading volume for a particular asset pair. We exclude burn events
with zero liquidity withdrawal in both base and quote assets, as traders use them solely to collect
fees without altering their liquidity position.
These basic screens give us a baseline sample of 274 liquidity pools covering 242 asset pairs, with
combined daily dollar volume of $1.12 billion and total value locked (i.e., aggregate liquidity supply)
of $2.53 billion as of July 15, 2023. We capture 24,202,803 interactions with liquidity pool smart
contracts (accounting for 86.04% of the entire universe of trades and liquidity updates). Trading
and liquidity provision on Uniswap is heavily concentrated: the five largest pairs (USDC-WETH,
WETH-USDT, USDC-USDT, WBTC-WETH, and DAI-USDC) account on average for 86% of
trading volume and 63% of supplied liquidity.6

3.2 Liquidity fragmentation patterns

For 32 out of the 242 asset pairs in our baseline sample, liquidity supply is fragmented across two
pools with different fees – either with 1 and 5 bps fees (5 pairs), 5 and 30 bps fees (6 pairs), or 30
and 100 bps fees (21 pairs).7 Despite being fewer in number, fragmented pairs are economically
important: they account on average for 95% of the capital committed to Uniswap v3 and for
93% of its dollar trading volume. All major token pairs such as WETH-USDC, WETH-USDT, or
WBTC-WETH trade on fragmented pools.
For each fragmented liquidity pair, we label the low and the high fee liquidity pool to facilitate
analysis across assets. For example, the low and high liquidity fees for USDC-WETH are 5 and 30
6
WETH and WBTC stand for “wrapped” Bitcoin and Ether. Plain vanilla Bitcon and Ether are not compliant with
the ERC-20 standard for tokens, and therefore cannot be directly used on decentralized exchanges’ smart contracts.
USDC (USD Coin), USDT (Tether), and DAI are stablecoins meant to closely track the US dollar.
7
In some cases, more than two pools are created for a pair – e.g., for USDC-WETH there are four pools with 1, 5,
30, and 100 bps liquidity fees. In all cases however, two pools heavily dominate the others: As described in Section 3.1
we filter out small pools with less than 1% volume share or less than $100,000 liquidity deposits.

15
bps, respectively, but only 1 and 5 bps for a lower volatility pair such as USDC-USDT. We refer to
non-fragmented pools as single (i.e., the unique pool for an asset pair).
We aggregate all interactions with Uniswap smart contracts into a panel across days and liquidity
pools. To compute the end-of-day pool size, we account for all changes in token balances, across all
price ranges. There are three possible interactions: A deposit or “mint” adds tokens to the pool,
a withdrawal or “burn” removes tokens, whereas a trade or “swap” adds one token and removes
the other. We track these changes across to obtain daily variation in the quantity of tokens on
each pool. We obtain dollar values for the end-of-day liquidity pool sizes, intraday trade volumes,
and liquidity events from the Uniswap V3 Subgraph. To determine a token’s price in dollars, the
Subgraph searches for the most liquid path on Uniswap pools to establish the token’s price in Ether
and subsequently converts the Ether price to US dollars.
Table 1 reports summary statistics across pools with different fee levels. High-fee pools attract
on average 58% of total liquidity supply, significantly more than their low-fee counterparts ($46.50
million and $33.78 million, respectively), but only capture 20.74% percent of the trading volume
(computed as 8,071.24/(8,071.24+30,848.79) from the first column of Table 1). Consistent with our
theoretical predictions, low-fee pools attract five times as many trades as high-fee competitors (610
versus 110 average trade count per day). At the same time, the average trade on a high-fee pool is
twice as large ($14,490) than on a low-fee pool ($6,340).
The distribution of mint sizes is heavily skewed to the right, with 6.6% of deposits exceeding
$1 million. There are large differences across pools – the median LP deposit on the low-fee pool is
$15,680, twice as much as the median deposit on the high-fee pool ($7,430). At the same time, the
number of liquidity providers on high-fee pools is 51% higher than on low-fee pools (10.08 unique
addresses per day on high-fee pools versus only 6.68 unique address on high-fee pools).
One concern with measuring average mint size is just-in-time liquidity provision (JIT). JIT
liquidity providers submit very large and short-lived deposits to the pool to dilute competitors on
an incoming large trade; they immediately withdraw the balance in the same block after executing
the trade. In our sample, JIT liquidity provision is not economically significant, accounting for less
than 1% of aggregate trading volume. However, it has the potential to skew mint sizes to the right,
particularly in low-fee pools, without providing liquidity to the market at large. We address this
issue by (i) filtering out JIT mints using the algorithm in Appendix D and (ii) taking the median
liquidity mint size at day-pool level rather than the mean.
Further, we follow Augustin, Chen-Zhang, and Shin (2022) to compute the daily liquidity fee
yield as the product between pool’s fee tier and the ratio between trading volume and the lagged
total value locked (TVL). That is,

Volumei,t
Liquidity yield = liquidity feei × , (21)
TVLi,t−1

16
 Table 1: Descriptive statistics
This table reports descriptive statistics for variables used in the empirical analysis. Pool size is defined as the total
value locked in the pool’s smart contract at the end of each day. We compute the balance on day t as follows: we
take the balance at day t − 1 and add (subtract) liquidity deposits (withdrawals) on day t, as well as accounting for
token balance changes due to trades. The liquidity balance on the first day of the pool is taken to be zero. End of
day balances are finally converted to US dollars. Daily volume is computed as the sum of US dollar volume for all
trades in a given pool and day. Liquidity share (Volume share) is computed as the ratio between a pool size (trading
volume) for a given fee level and the aggregate size of all pools (trading volumes) for the same pair in a given day.
Trade size and Mint size are the median trade and liquidity deposit size on a given pool and day, denominated in US
dollars. Trade count represents the number of trades in a given pool and day. LP wallets counts the unique number of
wallet addresses interacting with a given pool in a day. The liquidity yield is computed as the ratio between the daily
trading volume and end-of-day TVL, multiplied by the fee tier. The price range for every mint is computed as the
difference between the top and bottom of the range, normalized by the range midpoint – a measure that naturally lies
between zero and two. The impermanent loss is computed as in Heimbach, Schertenleib, and Wattenhofer (2022) for a
position in the range of 95% to 105% of the current pool price, with a forward-looking horizon of one hour. Finally,
mint-to-burn and burn-to-mint times are defined as the time between a mint (burn) and a subsequent burn (mint) by
the same address in the same pool, measured in hours. Mint-to-burn and burn-to-mint are recorded on the day of the
final interaction with the pool.

Statistic Pool fee Mean Median St. Dev. Pctl(25) Pctl(75) N

Pool size ($M) Low 33.78 2.05 96.91 0.30 14.12 20,151
High 46.50 3.85 95.73 1.43 27.51 20,151
Single 3.89 0.84 13.56 0.26 2.62 130,767
Liquidity share (%) Low 39.52 35.52 32.53 7.37 72.16 20,151
High 60.48 64.48 32.53 27.84 92.63 20,151
Daily volume ($000) Low 30,848.79 619.77 118,908.80 6.18 5,697.30 20,151
High 8,071.24 114.96 36,777.38 7.83 1,882.12 20,151
Single 915.73 36.07 6,059.78 1.93 277.00 130,767
Volume share Low 66.51 88.38 38.50 29.43 98.48 18,001
High 42.20 23.83 41.18 3.19 95.03 18,058
Trade size ($000) Low 6.34 2.20 13.36 0.61 6.03 18,001
High 14.49 2.76 33.19 0.82 10.48 18,060
Single 4.12 1.32 11.03 0.45 3.79 113,362
Mint size ($000) Low 820.84 15.68 13,114.83 3.78 58.98 10,640
High 1,001.10 7.43 13,807.10 1.55 30.52 10,370
Single 96.97 6.93 622.12 1.42 30.39 45,300
Trade count Low 610.61 95 1,518.52 12 414 20,151
High 110.59 26 490.29 8 89 20,151
Single 63.94 19 194.03 4 55 130,767
LP wallets Low 6.68 1 16.01 0 6 20,151
High 10.08 1 37.79 0 5 20,151
Single 1.57 1.17 1.19 1.00 1.85 55,580
Liquidity yield (bps) Low 11.72 2.58 56.31 0.16 9.08 20,122
High 9.69 1.65 51.44 0.15 6.40 20,130
Single 17.90 1.94 90.18 0.18 8.58 130,433
Price range Low 0.39 0.30 0.37 0.13 0.56 11,866
High 0.61 0.54 0.44 0.32 0.84 12,195
Single 0.68 0.58 0.52 0.27 1.02 55,580
Impermanent loss (bps) Low 8.46 1.84 27.88 0.06 7.23 20,118
High 7.37 1.33 27.21 0.05 5.93 20,132
Single 17.20 2.44 71.34 0.17 11.37 130,340
Mint-to-burn (hrs) Low 450.40 59.82 1,341.67 19.70 243.83 10,186
High 952.14 165.61 2,076.42 39.66 711.67 9,979
Single 760.26 126.64 1,778.62 27.01 563.50 39,735
Burn-to-mint (hrs) Low 105.29 0.20 521.62 0.08 5.63 8,279
High 224.26 0.32 941.31 0.10 27.78 7,289
Single 177.74 0.23 803.40 0.07 20.64 27,477
17
for pool i and day t. The average daily yield is slightly higher on low-fee pools, at 11.72 basis points,
compared to 9.69 basis points on high-fee pools.
A salient observation in Table 1 is that non-fragmented pairs (“single” pools) are significantly
smaller – on average less than 10% of the pool size and trading volume of fragmented pairs. Average
trade and mint sizes are correspondingly lower as well. The evidence suggests that pairs for which
there is significant trading interest, and therefore potentially a broader cross-section of potential
liquidity providers, are more likely to become fragmented.
Figure 5 plots the distributions of our empirical measures across low- and high-fee liquidity pools.
It suggest a sharp segmentation of liquidity supply and trading across pools. High-fee pools attract
smaller liquidity providers by mint size, and end up with a larger aggregate size than their low-fee
counterparts. Trading volume is similarly segmented: most small value trades are executed on the
cheaper low-fee pools, making up the majority of daily volume for a given pair. High-value trades,
of which there are fewer, are more likely to (also) execute on high-fee pools.
Our theoretical framework in Section 2 implies that liquidity suppliers manage their positions
more actively in the low- than the high-fee pool. Figure 6 provides suggestive evidence for liquidity
cycles of different lengths in the cross-section of pools. Liquidity on decentralized exchanges is
significantly more passive than on traditional equity markets. That is, liquidity providers do not
often manage their positions at high frequencies. The median time from a mint (deposit) to a
subsequent burn (withdrawal) from the same wallet on the same pool ranges from 59.82 hours, or
2.49 days, on low-fee pools to 165.61 hours, or 6.9 days on high-fee pools.
When do LPs re-balance their positions? In 53% of cases, liquidity providers only withdraw
tokens from the pool when their position exits the price range that allows them to collect fees.
Concretely, LPs specified price range for liquidity provision does not straddle the most recent
reference price of the pool. The scenario mirrors a limit order market where a liquidity provider’s
outstanding limit orders are deep in the book, such that she doesn’t stand to earn the spread on
the marginal incoming trade. In this case, a rational market maker might want to cancel their
outstanding order and place a new one at the top of the book. This is exactly the pattern we
observe on Uniswap: the subsequent mint following a burn straddles the new price 77% of the time –
LPs reposition their liquidity around the current prices to keep earning fees on incoming trades.
Moreover, re-balancing is swift – the median time between a burn and a subsequent mint is just 12
minutes (0.20 hours).
The empirical pattern in Figure 6 echoes liquidity cycles as described in Section 2. Liquidity
providers deposit tokens in Uniswap pools and then patiently wait for days until incoming order
flow exhausts their position (i.e., posted liquidity no longer earns fees). Once this happens, LPs
quickly re-balance their position in a matter of minutes – by removing stale liquidity and adding
a new position around the current price. The cycle is longer on high-fee pools for which trading

18
 Figure 5: Liquidity supply on decentralized exchanges
This figure plots the empirical distributions of variables in the pool-day panel, across low and high fee pools (for
fragmented pairs) as well as single pools in pairs that are not fragmented. In each box plot, the median is marked as a
vertical line; the box extends to the quartiles of the data set, whereas the whiskers extend to an additional 1.5 times
the inter-quartile range.

(a) Pool size and trading volume

(b) Average liquidity mint and trade size

(c) Number of liquidity providers and trades

19
 Figure 6: Liquidity cycles on high- and low-fee pools
The top panel plots the distribution of liquidity cycle times from mint to subsequent burn (left) and from burn to
subsequent mint (right) for the same LP wallet address in the same pool. In each box plot, the median is marked as a
vertical line; the box extends to the quartiles of the data set, whereas the whiskers extend to an additional 1.5 times
the inter-quartile range. The bottom panel plots the probability that the LP position is out of range and therefore does
not earn fees. A position is considered to be “out of range” when the minimum and maximum prices at which the LP
is willing to provide liquidity do no straddle the current price on the pool. We plot the probability separately for low-
and high- fees, as well as conditional on whether the event is a burn (liquidity withdrawal) or mint (liquidity deposit).

volume is lower and liquidity takes longer to deplete.

Importantly, LPs do not seem to “race” to update liquidity upon information arrival as in Budish,
Cramton, and Shim (2015). First, they very rarely manage their position intraday. Second, LPs on
Uniswap typically do not remove in-range liquidity that stands to trade first against incoming order
flow and therefore bears the highest adverse selection risk. Our results are consistent with Capponi
and Jia (2021) who theoretically argue that LPs have low incentives to manage liquidity on news
arrival, as well as with Capponi, Jia, and Yu (2022) who find no evidence of traders racing to trade
on information on Uniswap v2.

Measuring gas prices. Each interaction with smart contracts on the Ethereum blockchain requires
computational resources, measured in units of “gas.” Upon submitting a mint or burn transaction to
the decentralized exchange, each liquidity provider specifies their willingness to pay per unit of gas,
that is they bid a “gas price.” Traders are likely to bid higher prices for more complex transactions

20
 Figure 7: Gas costs for Uniswap v3 mint/burn transactions
The figure illustrates the daily average gas cost on mint/burn transactions in Uniswap v3 pools. The gas cost is
computed as the average of the lowest 1000 user gas bids for mint and burn interactions on each day, across all liquidity
pools in the benchmark sample.

or if they require a faster execution. To generate a conservative daily benchmark for the gas price,
we compute the average of the lowest 1000 user gas bids for mint and burn interactions on day t,
across all liquidity pools in the benchmark sample.
Figure 7 showcases the significant fluctuation in gas costs for Uniswap liquidity transactions over
time. Gas costs denominated in USD are influenced by two primary factors: network congestion,
which leads to variations in gas prices measured in Ether, and the fluctuation of Ether’s value
relative to the US dollar. On a monthly average, gas costs peaked at above US$100 in November
2021 and have since plummeted to around US$6 from the second half of 2022, albeit with occasional
spikes.

4 Empirical results

4.1 Liquidity supply on high- and low-fee pools

To formally test the model predictions and quantify the differences in liquidity supply across
fragmented pools, we build a panel data set for the 32 fragmented pairs in our sample where the
unit of observation is pool-day. We estimate linear regressions of liquidity and volume measures on
liquidity fees and gas costs:
X
yijt = α + β0 dlow-fee, ij + β1 GasPricejt + β2 GasPricejt × dlow-fee, ij + βk Controlsijt + θj + δw + εijt ,
(22)

21
where y is a variable of interest, i indexes liquidity pools, j runs over asset pairs, and t and w
indicates days and weeks, respectively. The dummy dlow-fee, ij takes the value one for the pool with
the lowest fee in pair j and zero else.
Further, our set of controls includes pair and week fixed effects, the log aggregate trading volume
and log liquidity supply (i.e., total value locked) for day t across all pools i. Volume and liquidity are
measured in US dollars. We also control for daily return volatility, computed as the range between
the daily high and low prices for a given pair j (following Alizadeh, Brandt, and Diebold, 2002):

Highjt
 
1
Volatilityjt = √ log . (23)
2 log 2 Lowjt

To measure volatility for fragmented pairs that actively trade in multiple pools, we select the pool
with the highest trading volume for a given day.
Consistent with Figure 5, most of the capital deployed to provide liquidity for a given pair is
locked in high-fee pools. At the same time, low-fee pools attract much larger trading volume. Models
(1) and (5) show that the average low-fee pool attracts 39.5% of liquidity supply for the average pair
(that is, equal to (100−20.92)/2) while it executes 62% (i.e., (100+24.62)/2) of the total trading volume.
At a first glance, it would seem that a majority of capital on decentralized exchanges is inefficiently
deployed in pools with low execution probability. We will show that, in line with our model, the
difference is driven by heterogeneous liquidity cycles across pools, leading to the formation of LP
clienteles.
The regression results in Table 2 support Prediction ??, stating that market share differences
between pools are linked to variation in fixed transaction costs on the blockchain. A one-standard
deviation increase in gas prices leads to a 4.63 percentage point increase in the high-fee liquidity
share. The results suggests that blockchain transaction costs have an economically meaningful and
statistically significant impact on liquidity fragmentation. In line with the theoretical model in
Section 2, a jump in gas prices leads to a reshuffling of liquidity supply from low- to high-fee pools.
Evidence suggests that a higher gas price leads to a 6.52% lower volume share for the low-fee
pool. This outcome is natural, as the incoming order flow is optimally routed to the high-fee pool,
following the liquidity providers.
What drives the market share gap across fragmented pools? In Table 3 we document stark
differences between the characteristics of individual orders supplying or demanding liquidity on
pools with low and high fees. On the liquidity supply side, model (1) in Table 3 shows that the
average liquidity mint is 107.5% larger on low-fee pools, which supports Prediction ?? of the model.8
At the same time, there are 3.40 fewer unique wallets (Model 5) providing liquidity on the low-fee
8
Since all dependent variables are measured in natural logs, the marginal impact of a dummy coefficient β is
computed eβ − 1 × 100 percent.

22
 Table 2: Liquidity pool market shares and gas prices
This table reports the coefficients of the following regression:
X
MarketShareijt = α + β0 dlow-fee, ij + β1 GasPricejt + β2 GasPricejt × dlow-fee, ij + βk Controlsijt + θj + εijt

where the dependent variable is the liquidity or trading volume market share for pool i in asset pair j on day t.
dlow-fee, ij is a dummy that takes the value one for the pool with the lowest fee in pair j and zero else. GasPricejt
is the average of the lowest 100 bids on liquidity provision events across all pairs on day t, standardized to have a
zero mean and unit variance. Volume is the natural logarithm of the sum of all swap amounts on day t, expressed in
thousands of US dollars. Total value locked is the natural logarithm of the total value locked on Uniswap v3 pools
on day t, expressed in millions of [Link] is computed as the daily range between high and low prices on
the most active pool for a given [Link] regressions include pair and week fixed-effects. Robust standard errors in
parenthesis are clustered by week and ***, **, and * denote the statistical significance at the 1, 5, and 10% level,
respectively. The sample period is from May 4, 2021 to July 15, 2023.

Liquidity market share (%) Volume market share (%)

(1) (2) (3) (4) (5) (6) (7) (8)
dlow-fee -20.92*** -20.92*** -20.92*** -20.94*** 24.62*** 24.63*** 24.62*** 24.71***
(-27.42) (-27.41) (-27.42) (-23.95) (20.55) (20.56) (20.55) (18.54)
Gas price × dlow-fee -4.63*** -4.62*** -4.63*** -6.52*** -6.52*** -6.52***
(-7.32) (-7.32) (-7.32) (-5.92) (-5.92) (-5.92)
Gas price 2.31*** 2.31*** 2.31*** 3.63*** 3.61*** 3.61***
(7.32) (7.32) (7.32) (7.33) (7.30) (7.26)
Volume 0.00 0.00 0.00 0.00 -0.19** -0.20** -0.19** -0.12
(0.65) (1.33) (0.65) (0.66) (-2.54) (-2.61) (-2.50) (-1.56)
Total value locked -0.00 -0.00 -0.00 0.58 0.58 0.44
(-0.58) (-0.06) (-0.64) (1.44) (1.44) (1.10)
Volatility -0.29 -0.29 -0.28 -1.15*** -1.15*** -1.13**
(-0.90) (-0.90) (-0.82) (-2.74) (-2.74) (-2.56)
Constant 60.45*** 60.46*** 60.45*** 60.46*** 41.96*** 41.99*** 41.96*** 41.96***
(158.00) (158.46) (158.00) (137.54) (69.99) (70.22) (70.02) (62.81)
Pair FE Yes Yes Yes Yes Yes Yes Yes Yes
Week FE Yes Yes Yes Yes Yes Yes Yes Yes
Observations 40,288 40,288 40,288 40,288 36,059 36,059 36,059 36,059
R-squared 0.10 0.10 0.10 0.09 0.13 0.13 0.13 0.12
Robust t-statistics in parentheses. Standard errors are clustered at week level. *** p<0.01, ** p<0.05, * p<0.1

23
pool – that is, a 34% relative difference between high- and low-fee pools.
On the liquidity demand side, trades on the low-fee pool are 25.91% smaller (Model 2), consistent
with Prediction ??. However, the low-fee pool executes almost three times the number of trades
(i.e., trade count is 177% higher from Model 4) and has 143% higher volume than the high-fee pool
(Model 3). On average, liquidity providers on low-fee pools earn 2.03 basis points higher revenue
than their counterparts on high-fee pools (Model 6), indicating significant positive returns resulting
from economies of scale.
Our findings (Model 7) indicate that liquidity providers on low-fee pools select price ranges that
are 30% (=0.18/0.59) narrower when minting liquidity compared to those on high-fee pools. This
pattern aligns with the capability of large LPs to adjust their liquidity positions frequently, enabling
more efficient capital concentration. Similarly, Caparros, Chaudhary, and Klein (2023) report a
higher concentration of liquidity in pools on alternative blockchains like Polygon, known for lower
transaction costs than Ethereum.
The results point to an asymmetric match between liquidity supply and demand across pools.
On low-fee pools, a few LPs provide large chunks of liquidity for the vast majority of incoming small
trades. Conversely, on high-fee pools there is a sizeable mass of small liquidity providers that mostly
trade against a few large incoming trades.
How does variation in fixed transaction costs impact the gap between individual order size across
pools? We find that increasing the gas price by one standard deviation leads to higher liquidity
deposits on both the low- and the high-fee pools (14.2% and 30.1% higher, respectively).9 The
result supports Prediction ?? of the model. Our theoretical framework implies that a larger gas
price leads to some (marginal) LPs switching from the low- to the high-fee pool. The switching
LPs have low capital endowments relative to their low-fee pool peers, but higher than LPs on the
high-fee pool. Therefore, the gas-driven reshuffle of liquidity leads to a higher average endowment
on both high- and low-fee pools. Consistent with the model, a higher gas price leads to fewer active
liquidity providers, particularly on low-fee pools. Specifically, a one-standard increase in gas costs
leads to a significant decrease in the number of LP wallets interacting daily with low- and high-fee
pools, respectively (Model 5).
While a higher gas price is correlated with a shift in liquidity supply, it has a muted impact on
liquidity demand on low-fee pools. A higher gas cost is associated with 6% larger trades (Model 2),
likely as traders aim to achieve better economies of scale. At the same time, the number of trades
on the low-fee pool drops by 5.1% (Model 4) – since small traders might be driven out of the market.
The net of gas prices effect on aggregate volume on the low-fee pool is small and not statistically
significant (Model 3). The result matches our model assumption that the aggregate order flow on
9
The relative effects are computed as 0.37/(1.88+0.73) = 13.8% for low pools and 0.58/1.88 = 30.85% for high-fee pools,
respectively.

24
 Table 3: Fragmentation and order flow characteristics
This table reports the coefficients of the following regression:
X
yijt = α + β0 dlow-fee, ij + β1 GasPricejt dlow-fee, ij + β2 GasPricejt × dhigh-fee, ij + βk Controlsijt + θj + εijt

where the dependent variable yijt can be (i) the log median mint size, (ii) the log median trade size, (iii) the log
trading volume, (iv) the log trade count log(1 + #trades), (v) count of unique LP wallets interacting with a pool in a
given day, (vi) the liquidity yield in bps for pool i in asset j on day t, computed as in equation (21), and (vii) the
average liquidity mint price range for pool i in asset j on day t. Price range is computed as the difference between the
top and bottom of the range, normalized by the range midpoint – a measure that naturally lies between zero and
two. dlow-fee, ij is a dummy that takes the value one for the pool with the lowest fee in pair j and zero else. dhigh-fee, ij
is defined as 1 − dlow-fee, ij . GasPricejt is the average of the lowest 100 bids on liquidity provision events across all
pairs on day t, standardized to have a zero mean and unit variance. Volume is the natural logarithm of the sum of
all swap amounts on day t, expressed in thousands of US dollars. Total value locked is the natural logarithm of the
total value locked on Uniswap v3 pools on day t, expressed in millions of dollars. Volatility is computed as the daily
range between high and low prices on the most active pool for a given pair. All regressions include pair and week
fixed-effects. Robust standard errors in parenthesis are clustered by week and ***, **, and * denote the statistical
significance at the 1, 5, and 10% level, respectively. The sample period is from May 4, 2021 to July 15, 2023.

Mint size Trade size Volume # Trades # Wallets Liquidity yield Price range
(1) (2) (3) (4) (5) (6) (7)
dlow-fee 0.73*** -0.30*** 0.89*** 1.02*** -3.40*** 2.03*** -0.18***
(12.27) (-10.05) (14.23) (32.95) (-5.00) (3.60) (-41.84)
Gas price × dlow-fee 0.37*** 0.08*** -0.03 -0.22*** -3.00*** 3.57** -0.00
(4.96) (3.75) (-0.95) (-7.29) (-3.43) (2.30) (-0.47)
Gas price × dhigh-fee 0.58*** 0.17*** 0.24*** 0.07** -2.89*** 5.57*** -0.03***
(7.52) (8.81) (5.95) (2.46) (-3.15) (2.83) (-4.65)
Volume 0.37*** 0.16*** 0.43*** 0.20*** 1.22*** 1.01 -0.01**
(8.68) (21.38) (15.27) (13.85) (6.56) (0.81) (-2.56)
Total value locked -0.16 0.11*** 0.23** -0.01 -1.86 -13.42 -0.02
(-1.30) (3.54) (1.99) (-0.18) (-0.99) (-1.09) (-0.99)
Volatility -0.04 -0.01 -0.07 0.01 -0.09 1.18** 0.02***
(-1.11) (-1.34) (-1.38) (0.88) (-1.03) (2.21) (3.98)
Constant 1.88*** 1.64*** 5.27*** 3.26*** 10.12*** 10.01*** 0.59***
(58.27) (111.47) (168.58) (209.84) (28.65) (26.04) (184.91)

Pair FE Yes Yes Yes Yes Yes Yes Yes

Week FE Yes Yes Yes Yes Yes Yes Yes
Observations 21,000 36,059 36,059 40,288 40,288 40,252 24,058
R-squared 0.26 0.53 0.55 0.52 0.37 0.09 0.42
Robust t-statistics in parentheses. Standard errors are clustered at week level.
*** p<0.01, ** p<0.05, * p<0.1

25
low-fee pool is not sensitive to gas prices.10
On the high-fee pool, a higher gas price is also associated with a higher trade size, but a much
smaller relative increase in the number of trades. The implication is that large-size traders who
typically route orders to the high-fee pool are unlikely to be deterred by an increase in fixed costs,
while they do adjust quantities to achieve better economies of scale. In the context of our model,
the large LT arrival rate λ does not depend on the gas price Γ.
In Table 4, we shift the analysis from individual orders to aggregate daily liquidity flows to
Uniswap pools. We find that higher gas prices lead to a decrease in liquidity inflows, but only on the
low fee pools. A one standard deviation increase in gas prices leads to a 29.5% drop in new liquidity
deposits by volume (Model 1) and an 6.02% drop in probability of having at least one mint (Model
4) on the low-fee pool. However, the slow-down in liquidity inflows is less evident in high fee pools.
While an increase in gas prices reduce the probability of liquidity inflows by 1.43%, it actually leads
to a 11.6% increase in the daily dollar inflow to the pool. Together with the result in Table 3 that
the size of individual mints increases with gas prices, our evidence is consistent with the model
implication that higher fixed transaction costs change the composition of liquidity supply on the
high-fee pool, with small LP being substituted by larger LPs switching over from the low-fee pool.

4.2 Liquidity cycles on high- and low-fee pools

Next, we test Predictions ?? and ?? on the duration of liquidity cycles on fragmented pools. Since
the descriptive statistics in Table 1 suggest that LPs manage their positions over multiple days, we
cannot accurately measure liquidity cycles in a pool-day panel. Instead, we use intraday data on
liquidity events (either mints or burns) to measure the duration between two consecutive opposite-
sign interactions by the same Ethereum wallet with a liquidity pool: either a mint followed by a
burn, or vice-versa.
To complement our previous analysis, we additionally control for whether each liquidity position
is out of range (i.e., the price range set by the LP does not straddle the current price and therefore
the LP does not earn fees). We further introduce wallet fixed effects to soak up variation in reaction
times across traders.
Table 5 presents the results. Liquidity updates on decentralized exchanges are very infrequent,
as times elapsed between consecutive interactions are measured in days or even weeks. In line
with Prediction ??, we find evidence for shorter liquidity cycles on low-fee pools. The average time
between consecutive mint and burn orders is 20.06% shorter on the low-fee pool (from Model 1, the
relative difference is 99.74 hours/497.18 hours).
10
Formally, one could extend the model to assume that small LTs arrive at the market at rate θ̃ (Γ) dt and demand
f (Γ) units each, where θ̃ (Γ) decreases in Γ and f (Γ) increases in Γ such that θ̃ (Γ) f (Γ) = θ.

26
 Table 4: Liquidity flows and gas costs on fragmented pools
This table reports the coefficients of the following regression:
X
yijt = α + β0 dlow-fee, ij + β1 GasPricejt dlow-fee, ij + β2 GasPricejt × dhigh-fee, ij + βk Controlsijt + θj + εijt

where the dependent variable yijt can be (i) the aggregate dollar value of mints (in logs), or (vi) a dummy variable
taking value one hundred if there is at least one mint on liquidity pool i in asset j on day t. dlow-fee, ij is a dummy
that takes the value one for the pool with the lowest fee in pair j and zero else. dhigh-fee, ij is defined as 1 − dlow-fee, ij .
GasPricejt is the average of the lowest 100 bids on liquidity provision events across all pairs on day t, standardized
to have a zero mean and unit variance. Volume is the natural logarithm of the sum of all swap amounts on day t,
expressed in thousands of US dollars. Total value locked is the natural logarithm of the total value locked on Uniswap
v3 pools on day t, expressed in millions of [Link] is computed as the daily range between high and low
prices on the most active pool for a given [Link] regressions include pair and week fixed-effects. Robust standard
errors in parenthesis are clustered by week and ***, **, and * denote the statistical significance at the 1, 5, and 10%
level, respectively. The sample period is from May 4, 2021 to July 15, 2023.

Daily mints (log US$) Prob (at least one mint)

(1) (2) (3) (4) (5) (6)
dlow-fee 0.43*** 0.43*** 0.43*** 1.38* 1.37* 1.38*
(6.07) (6.07) (6.07) (1.71) (1.71) (1.71)
Gas price × dlow-fee -0.35*** -0.35*** -0.46*** -6.02*** -6.01*** -4.58***
(-8.50) (-8.50) (-7.14) (-9.13) (-9.13) (-6.76)
Gas price × dhigh-fee 0.11** 0.11** -1.43** -1.43**
(2.15) (2.15) (-2.57) (-2.57)
Volume 0.26*** 0.26*** 0.26*** 0.96*** 0.96*** 0.96***
(14.78) (14.77) (14.78) (3.93) (3.93) (3.93)
Total value locked -0.07 -0.07 -0.07 1.47 1.47 1.47
(-0.78) (-0.78) (-0.78) (1.01) (1.00) (1.01)
Volatility -0.01 -0.01 0.26 0.26
(-0.68) (-0.68) (0.59) (0.59)
Gas price 0.11** -1.43**
(2.15) (-2.57)
Constant 2.61*** 2.61*** 2.61*** 51.44*** 51.43*** 51.44***
(73.46) (73.49) (73.46) (126.67) (127.28) (126.67)

Pair FE Yes Yes Yes Yes Yes Yes

Week FE Yes Yes Yes Yes Yes Yes
Observations 40,288 40,288 40,288 40,288 40,288 40,288
R-squared 0.47 0.47 0.47 0.28 0.28 0.28
Robust t-statistics in parentheses. Standard errors are clustered at week level.
*** p<0.01, ** p<0.05, * p<0.1

27
 Liquidity cycles are in part driven by fixed Blockchain transaction costs. A one standard
deviation increase in gas prices speeds up the liquidity cycle on low-fee pools by a further 15.80
hours, a result which supports the intuition behind Prediction ??. When the gas price spikes, the
liquidity supply on the low-fee pool decreases at a higher rate than the liquidity demand. As a
result, liquidity deposits deplete faster (i.e., they move outside the fee-earning range), triggering the
need for more frequent updates. At the same time, a higher gas fee speeds up the liquidity cycle on
the high-fee pool as well. The result is consistent with small liquidity providers on the high-fee pool
being crowded out by the high fixed costs, leading to a lower liquidity supply.
The reaction time of a trader may depend on capital constraints, network congestion, and other
confounding factors that can correlate with gas prices. As a robustness check, we repeat the analysis
above with burn-to-mint times as the dependent variables. The burn-to-mint time measures the
speed at which LPs deposit liquidity at updated prices after removing (out-of-range) positions, and
should not depend on the rate at which liquidity is consumed. Consistent with the theory, we find
no significant relationship between gas price and the burn-to-mint duration (Models 4 through 6).

4.3 Adverse selection costs across low- and high-fee pools

In our model, the first-order friction driving liquidity fragmentation is the fixed gas cost, which leads
to heterogeneous rebalancing costs for small and large market makers respectively. Thus, pool fees
in our setup simply offset liquidity providers’ costs to manage their position. An alternative channel
driving fragmentation might be adverse selection: The rationale is that liquidity providers gravitate
towards high-fee pools as they provide superior protection against informed traders (following the
arguments in Milionis, Moallemi, and Roughgarden, 2023).
For example, consider a token pair traded in the 5 (low) and 30 (high) basis point fee pools.
If an arbitrage opportunity is worth 50 basis points and the relative gas cost is 15 basis points,
arbitrageurs would optimally execute swaps on both pools since the trade value surpasses total
cost, but obtain lower profits on the high-fee pool. However, if gas cost rises to 25 basis points,
arbitrageurs would only trade in the low fee pool. Thus, large gas costs might make the high-fee
pool more appealing to liquidity providers, as it offers added protection against adverse selection.
The mechanism could provide an alternative explanation for the result in Table 2 showing that
market makers shift to high-fee pools when gas prices rise. In this section, we investigate differences
in adverse selection across low- and high-fee pools, as well as the role of asymmetric information
costs in driving liquidity fragmentation.
Our main metric for informational costs is the loss-versus-rebalancing (LVR), as defined in
Milinois, Moallemi, Roughgarden, and Zhang (2023). The measure is equivalent to the adverse
selection component of the bid-ask spread in equity markets. To calculate it, for each swap j

28
 Table 5: Liquidity cycles on fragmented pools
This table reports the coefficients of the following regression:
X
yijtk = α + β0 dlow-fee, ij + β1 GasPricejt dlow-fee, ij + β2 GasPricejt × dhigh-fee, ij + βk Controlsijt + θj + εijt

where the dependent variable yijt can be (i) the mint-to-burn time, (ii) the burn-to-mint time, measured in hours, for
a transaction initiated by wallet k on day t and pool i trading asset j. The mint-to-burn and burn-to-mint times
are computed for consecutive interactions of the same wallet address with the liquidity pool. dlow-fee, ij is a dummy
that takes the value one for the pool with the lowest fee in pair j and zero else. dhigh-fee, ij is defined as 1 − dlow-fee, ij .
GasPricejt is the average of the lowest 100 bids on liquidity provision events across all pairs on day t, standardized
to have a zero mean and unit variance. Volume is the natural logarithm of the sum of all swap amounts on day t,
expressed in thousands of US dollars. Total value locked is the natural logarithm of the total value locked on Uniswap
v3 pools on day t, expressed in millions of dollars. Volatility is computed as the daily range between high and low
prices on the most active pool for a given pair. Position out-of-range is a dummy taking value one if the position
being burned or minted is out of range, that is if the price range selected by the LP does not straddle the current pool
price. All variables are measured as of the time of the second leg of the cycle (i.e., the burn of a mint-burn cycle). All
regressions include pair, week, and trader wallet fixed-effects. Robust standard errors in parenthesis are clustered by
day and ***, **, and * denote the statistical significance at the 1, 5, and 10% level, respectively. The sample period is
from May 4, 2021 to July 15, 2023.

Mint-burn time Burn-mint time

(1) (2) (3) (4) (5) (6)
dlow-fee -99.74*** -100.17*** -104.09*** -157.95*** -159.71*** -159.69***
(-8.86) (-8.94) (-9.22) (-10.59) (-10.81) (-10.80)
Gas price × dlow-fee -16.65** -15.41* -15.80** -11.29 2.95 2.94
(-2.13) (-1.98) (-2.02) (-1.65) (0.40) (0.39)
Gas price × dhigh-fee -14.44** -13.42* -13.98* -10.52* 1.96 1.95
(-2.04) (-1.89) (-1.96) (-1.69) (0.32) (0.32)
Volume -5.87 -7.45 -24.84*** -24.82***
(-1.15) (-1.41) (-4.10) (-4.09)
Total value locked -53.17* -51.72* -12.71 -12.71
(-1.70) (-1.66) (-0.52) (-0.52)
Volatility -2.11*** -2.26*** -2.99*** -2.98***
(-2.75) (-2.86) (-3.36) (-3.36)
Position out-of-range 37.09*** -1.53
(6.43) (-0.22)
Constant 497.18*** 497.00*** 479.22*** 248.00*** 250.13*** 250.47***
(91.65) (90.60) (82.13) (29.91) (30.27) (30.32)

Pair FE Yes Yes Yes Yes Yes Yes

Week FE Yes Yes Yes Yes Yes Yes
Trader wallet FE Yes Yes Yes Yes Yes Yes
Observations 405,586 405,584 405,584 265,848 265,848 265,848
R-squared 0.82 0.82 0.82 0.37 0.37 0.37
Robust t-statistics in parentheses. Standard errors are clustered at week level.
*** p<0.01, ** p<0.05, * p<0.1

29
exchanging ∆xj for ∆yj in a pool with assets x and y, we use:

LVRj = dj × ∆xj (pswap,j − p′j ), (24)

where dj is one for a “buy” trade (∆xj < 0) and minus one for a “sell” trade (∆xj > 0). The effective
∆y
swap price is pswap,j = − ∆xjj , and p′j represents a benchmark price.
We use two benchmark prices p′j in our analysis. The first, p′j = p∆t=0
j , is the pool’s equilibrium
price immediately after a swap. The resulting LVR metric captures both temporary and permanent
price impact, driven by uninformed and informed trades, respectively, and represents an upper
bound for LP’s adverse selection cost.
The second benchmark is the liquidity-weighted average price across Uniswap v3 pools, measured
with a one-hour delay after the swap (p′j = p∆t=1h
j ). This approach assumes that any price deviations
caused by uninformed liquidity trades are corrected by arbitrageurs within an hour, as supported by
Lehar and Parlour (forthcoming). Thus, the LVR metric derived using this benchmark captures
only the permanent price impact, a more precise measure of adverse selection cost for liquidity
providers.11
To compute LVR for each day t and liquidity pool i, we aggregate the loss-versus-balancing for
each swap within a day. We subsequently winsorize our measures at the 0.5% and 99.5% quantiles
to remove extreme outliers. The resulting sum is normalized by dividing it by the total value locked
(TVL) in the pool at day’s end: P
j LVRj,i,t
LVRi,t = , (25)
TVLi,t
which ensures that the LVR metric is comparable across pools trading different token pairs.
We complement our analysis with the calculation of impermanent loss (IL), an additional metric
for assessing adverse selection costs. Impermanent loss is defined as the negative return from
providing liquidity compared to simply holding the assets outside the exchange and marking them
to market as prices change (see, for example, Aoyagi, 2020; Barbon and Ranaldo, 2021).
The key distinction between IL and loss-versus-rebalancing (LVR) measures lies in their assump-
tions about liquidity providers’ strategies (Milinois, Moallemi, Roughgarden, and Zhang, 2023).
While LVR assumes that providers actively re-balance their holdings by mirroring decentralized
exchange trades on centralized exchanges at the fundamental value to hedge market risk, IL is based
on a more passive approach where providers maintain their positions without active re-balancing.
11
Our methodology is equivalent to the one in Milinois, Moallemi, Roughgarden, and Zhang (2023) under two
assumptions. First, liquidity providers can re-balance their position following each swap. Second, we assume that
our two benchmarks for p′j , derived from decentralized exchange data, closely track the fundamental value of the
token. This perspective aligns with Han, Huang, and Zhong (2022), who also note that centralized exchange prices are
subject to manipulative practices such as wash trading. Further, our selection of benchmarks reflects the fact that our
sample includes several token pairs not traded on major centralized exchanges such as Binance.

30
Loss-versus-rebalancing is a function of the entire price path, reflecting constant rebalancing by
liquidity providers. In contrast, impermanent loss is determined solely by the initial and final prices
of the assets.
We obtain hourly liquidity snapshots from the Uniswap V3 Subgraph to calculate impermanent
loss for a theoretical symmetric liquidity position. This position is set within a price range of
1 
α p, αp , centered around the current pool price p, with α set to 1.05. We set a one-hour horizon
to measure changes in position value, aligning with the time horizon used for the LVR metric. In
Appendix E, we present the exact formulas for calculating impermanent loss on Uniswap V3, based
on the methodology described by Heimbach, Schertenleib, and Wattenhofer (2022).
Table 6 presents our empirical results. In line with Milionis, Moallemi, and Roughgarden (2023),
all price impact measures — that is, the immediate and one-hour horizon LVR and the impermanent
loss — are significantly larger in low-fee pools. This indicates that a higher liquidity fee indeed
acts as barrier to arbitrageurs. Specifically, the permanent price impact, measured by the one-hour
horizon LVR, is 6.85 basis points or 81% larger in low-fee pools. The total price impact, represented
by the after-swap LVR metric, is 3.5 times larger in low-fee pools. The wide gap between permanent
and total price impact highlights the substantially higher volume of uninformed trading in low-fee
pools. Further, an increase in gas prices leads to a 3.51 wider gap in total price impact but a
0.75 bps narrower gap in permanent price impact between the high- and low-fee pools. The result
suggests that a higher gas price primarily discourages uninformed traders, rather than arbitrageurs,
from trading on high-fee pools.
The findings in Table 2 reveal that a higher gas cost translates to a lower liquidity and volume
market share for low-fee pools. We attribute this effect primarily to the fixed costs associated with
re-balancing. To examine whether the impact of gas prices on market share is driven by adverse
selection, in Models (7) and (8) we add an interaction between our permanent price impact measure
(one-hour horizon LVR), a dummy variable for low-fee pools, and the standardized gas price. The
analysis reveals that adverse selection contributes only 8% (computed as 0.38/4.41+0.38) to the effect
of gas prices on liquidity market shares. Moreover, this coefficient is only marginally statistically
significant. A similar pattern is observed for trading volume shares in Model 8.
Finally, in Models (9) and (10) of Table 6 we explore whether various arbitrage frictions lead
to price discrepancies between high- and low-fee pools. For this purpose, we collect hourly price
data from Binance for the largest four pairs by trading volume: WBTC-WETH, USDC-WETH,
WETH-USDT, and USDT-USDC. We subsequently compute daily averages of hourly price deviations
between centralized and decentralized exchanges. The analysis reveals that the average hourly price
deviation across centralized and decentralized exchanges is 0.60%. Notably, there is no significant
difference in price deviations between low- and high-fee pools. The result suggests that arbitrage
activities remain efficient despite the differences in trading costs between these pools.

31
 Table 6: Adverse selection costs on high- and low-fee pools
This table presents regression results that analyze adverse selection costs in fragmented Uniswap v3 pools. For columns
(1) through (4), the dependent variable is loss-versus-rebalancing (LVR), as defined in equation (25). We use the
one-hour horizon benchmark (p∆t=1h
j ) in models (1) and (2) to measure permanent price impact, and the immediate,
same-block price benchmark (p∆t=0j ) in models (3) and (4) to measure total price impact. For columns (5) and (6),
the dependent variable is the impermanent loss for a symmetric liquidity position at ±5% centered around the current
pool price. The average impermanent loss is calculated for each day, based on Ethereum blocks mined within that day.
The impermanent loss computation uses a one-hour liquidity provider horizon, comparing current pool prices with
those one hour later. For columns (7) and (8), the dependent variables are the liquidity (TVL) and volume share of
the pool, measured in percent. Finally, in columns (9) and (10) the dependent variable is the absolute deviation of
the Uniswap pool price from Binance prices, sampled hourly, and measured in percent. dlow-fee, ij is a dummy that
takes the value one for the pool with the lowest fee in pair j and zero else. GasPricejt is the average of the lowest
100 bids on liquidity provision events across all pairs on day t, standardized to have a zero mean and unit variance.
Volume is the natural logarithm of the sum of all swap amounts on day t, expressed in thousands of US dollars. Total
value locked is the natural logarithm of the total value locked on Uniswap v3 pools on day t, expressed in millions of
dollars. When LVR is an explanatory variable, it is calculated using the one-hour ahead benchmark price. Volatility is
computed as the daily range between high and low prices on the most active pool for a given pair. All regressions
include pair and week fixed-effects. Robust standard errors in parenthesis are clustered by week, and ***, **, and *
denote the statistical significance at the 1, 5, and 10% level, respectively. The sample period is from May 4, 2021 to
July 15, 2023.

LVR (1h horizon) LVR (after swap) Impermanent loss TVL share Volume share CEX price deviation
Permanent price impact Total price impact
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
dlow-fee 6.39*** 6.39*** 29.78*** 29.67*** 1.08*** 1.13*** -21.00*** 24.43*** 0.06 -0.00
(16.57) (17.05) (14.86) (14.95) (5.72) (6.18) (-26.45) (19.79) (1.51) (-0.06)
Gas price × dlow-fee -0.75** 3.51** -0.01 -4.41*** -6.16*** 0.09
(-2.05) (2.10) (-0.05) (-6.52) (-5.64) (0.95)
Gas price × dlow-fee × LVR -0.38* -0.48 -0.10*
(-1.88) (-1.45) (-1.68)
Gas price 2.61*** 6.16** 3.71*** 1.53*** 2.82*** -0.09
(2.74) (2.53) (3.76) (4.73) (5.23) (-0.83)
Gas price × LVR 0.59*** 0.56*** 0.11**
(4.26) (2.80) (2.17)
Loss-versus-rebalancing (LVR) 0.41*** 0.67*** 0.11***
(2.72) (3.86) (6.05)
Volume 3.22*** 8.67*** 1.81*** -0.02 -0.30*** 0.18***
(8.15) (6.61) (6.22) (-0.56) (-3.18) (3.82)
Total value locked 0.53 -2.12 1.93 -0.13 0.46 -0.33***
(0.14) (-0.34) (0.74) (-0.52) (0.77) (-4.00)
Volatility 1.85*** 4.23*** 6.69** -0.36 -1.27*** 0.77***
(2.87) (3.17) (2.61) (-1.16) (-3.13) (3.16)
Constant 7.85*** 7.86*** 8.88*** 8.97*** 7.37*** 7.51*** 60.22*** 41.59*** 0.60*** 0.64***
(40.71) (36.89) (8.87) (8.88) (77.84) (61.32) (162.17) (71.02) (30.71) (38.64)

Pair FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Week FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Observations 40,302 40,288 40,302 40,288 40,250 40,248 40,288 36,059 5,207 5,207
R-squared 0.14 0.15 0.09 0.10 0.09 0.11 0.10 0.13 0.10 0.11
Robust t-statistics in parentheses. Standard errors are clustered at week level.
*** p<0.01, ** p<0.05, * p<0.1

32
 Figure 8: Price impact and price deviations across high- and low-fee pools
This figure plots the average total and permanent price impact, liquidity yield, and price deviation from centralized
exchanges across low and high fee pools for fragmented pairs.

Figure 8 graphically illustrates the result, contrasting permanent price impact measures against
the liquidity yield, as calculated in equation (21). Notably, before accounting for gas costs, we
observe that liquidity providers in low-fee pools experience losses on average: the average daily
permanent price impact in these pools is 14.21 basis points, which exceeds the fee revenue of 11.71
bps. In contrast, liquidity providers in high-fee pools approximately break even before considering
gas costs: the fee revenue amounts to 9.69 bps, which is slightly higher than the permanent price
impact of 7.85 bps. One should keep in mind, however, that the magnitude of losses from adverse
selection depends on the horizon at which we measure the loss-versus-rebalancing.

5 Conclusion

This paper argues that fixed costs associated with liquidity management drive a wedge between large
(institutional) and small (retail) market makers. In the context of blockchain-based decentralized
exchanges, the most evident fixed cost is represented by gas fees, where market makers compensate
miners and validators for transaction processing in proof-of-work, respectively in proof-of-stake
blockchains. Innovative solutions such as Proof of Stake (PoS) consensus algorithms and Layer
2 scaling aim to address the concern of network costs. However, even if gas fees were eliminated
entirely, individual retail traders still encounter disproportionate fixed costs in managing their
liquidity, such as the expenditure of time and effort.
Our paper highlights a trade-off between capital efficiency and the fixed costs of active manage-
ment. During the initial phase of decentralized exchanges, such as Uniswap V2, liquidity providers

33
were not able to set price limits, resulting in an even more passive liquidity supply and fewer
incentives for active position management. However, the mechanism implied that incoming trades
incurred significant price impact. To enhance the return on liquidity provision and reduce price
impact on incoming trades, modern decentralized exchanges (DEXs) have evolved to enable market
makers to fine-tune their liquidity positions, albeit at the expense of more active management.
We show, both theoretically and empirically, that fixed costs of liquidity management promote
market fragmentation across decentralized pools and generate clienteles of liquidity providers. Large
market makers, likely institutions and funds, have stronger economies of scale and can afford to
frequently manage their positions on very active low fee markets. On the other hand, smaller retail
liquidity providers become confined to high fee markets with scant activity, trading off a lower
execution probability against lower gas costs to update their positions. Since large liquidity providers
can churn their position at a faster pace, two thirds of the trading volume interacts with less than
half the capital locked on Uniswap V3.
Our findings indicate that substantial fixed costs can hinder the participation of small market
makers in the forefront of liquidity provision, where active order management is crucial. Instead,
smaller liquidity providers tend to operate on the market maker “fringe,” opting for a lower execution
probability in exchange for better prices. The results are particularly relevant the context of a
resurgence in retail trading activity and the ongoing evolution of technology that fosters market
structures aimed at enhancing broader access to financial markets.

References
Adams, Hayden, Noah Zinsmeister, Moody Salem, River Keefer, and Dan Robinson, 2021, Uniswap
v3 core, .

Alizadeh, Sassan, Michael W. Brandt, and Francis X. Diebold, 2002, Range-based estimation of
stochastic volatility models, The Journal of Finance 57, 1047–1091.

Aoyagi, Jun, 2020, Liquidity Provision by Automated Market Makers, Working paper.

, and Yuki Ito, 2021, Coexisting Exchange Platforms: Limit Order Books and Automated
Market Makers, Working paper.

Aspris, Angelo, Sean Foley, Jiri Svec, and Leqi Wang, 2021, Decentralized exchanges: The “wild
west” of cryptocurrency trading, International Review of Financial Analysis 77, 101845.

Augustin, Patrick, Roy Chen-Zhang, and Donghwa Shin, 2022, Reaching for yield in decentralized
financial markets, LawFin Working Paper No. 39. Available at SSRN: [Link]
abstract=4063228.

34
Barbon, Andrea, and Angelo Ranaldo, 2021, On the quality of cryptocurrency markets: Centralized
versus decentralized exchanges, .

Battalio, Robert, Shane A. Corwin, and Robert Jennings, 2016, Can brokers have it all? on the
relation between make-take fees and limit order execution quality, The Journal of Finance 71,
2193–2237.

Brolley, Michael, and David Cimon, 2020, Order flow segmentation, liquidity and price discovery:
The role of latency delays, Journal of Financial and Quantitative Analysis 55, 2555–2587.

Brolley, Michael, and Marius Zoican, 2023, On-demand fast trading on decentralized exchanges,
Finance Research Letters p. 103350.

Budish, Eric, Peter Cramton, and John Shim, 2015, The high-frequency trading arms race: Frequent
batch auctions as a market design response, Quarterly Journal of Economics 130, 1547–1621.

Caparros, Basile, Amit Chaudhary, and Olga Klein, 2023, Blockchain scaling and liquidity concen-
tration on decentralized exchanges, Available at SSRN: [Link]
or [Link]

Capponi, Agostino, and Ruizhe Jia, 2021, The Adoption of Blockchain-based Decentralized Ex-
changes, Working paper.

, and Shihao Yu, 2022, The Information Content of Blockchain Fees, Working paper.

, 2023, Price discovery on decentralized exchanges, .

Cimon, David A., 2021, Broker routing decisions in limit order markets, Journal of Financial
Markets 54, 100602.

Foucault, Thierry, Ohad Kadan, and Eugene Kandel, 2013, Liquidity cycles and make/take fees in
electronic markets, Journal of Finance 68, 299–341.

Foucault, Thierry, and Albert J. Menkveld, 2008, Competition for order flow and smart order
routing systems, The Journal of Finance 63, 119–158.

Han, Jianlei, Shiyang Huang, and Zhuo Zhong, 2022, Trust in defi: An empirical study of the
decentralized exchange, .

Hasbrouck, Joel, Thomas Rivera, and Fahad Saleh, 2022, The Need for Fees at a DEX: How Increases
in Fees Can Increase DEX Trading Volume, Working paper.

Heimbach, Lioba, Eric Schertenleib, and Roger Wattenhofer, 2022, Risks and Returns of Uniswap
V3 Liquidity Providers, Papers 2205.08904 [Link].

35
Korajczyk, Robert A, and Dermot Murphy, 2018, High-Frequency Market Making to Large Institu-
tional Trades, The Review of Financial Studies 32, 1034–1067.

Lehar, Alfred, and Christine Parlour, forthcoming, Decentralized Exchange: The Uniswap automated
market maker, Journal of Finance.

Milinois, Jason, Ciamac Moallemi, Tim Roughgarden, and Anthony Lee Zhang, 2023, Automated
market making and loss-versus-rebalancing, Manuscript.

Milionis, Jason, Ciamac C. Moallemi, and Tim Roughgarden, 2023, Automated market making and
arbitrage profits in the presence of fees, .

O’Hara, Maureen, and Mao Ye, 2011, Is market fragmentation harming market quality?, Journal of
Financial Economics 100, 459–474.

Pagano, Marco, 1989, Trading volume and asset liquidity, The Quarterly Journal of Economics 104,
255–274.

Pagnotta, Emiliano, and Thomas Philippon, 2018, Competing on speed, Econometrica 86, 1067–1115.

Park, Andreas, 2022, Conceptual flaws of decentralized automated market making, Working paper,
University of Toronto.

Wan, Xin, and Austin Adams, 2022, Just-in-time liquidity on the uniswap protocol, Working paper.

Zhu, Haoxiang, 2014, Do dark pools harm price discovery?, Review of Financial Studies 27, 747–789.

36
A Liquidity provision mechanism on Uniswap v3

In this appendix, we walk through a numerical example to illustrate the mechanism of liquidity
provision and trading on Uniswap V3 liquidity pools. To facilitate understanding, we highlight
the similarities and differences between the Uniswap mechanism and the familiar economics of a
traditional limit order book.
Let pc = 1500.62 be the current price of the ETH/USDT pair. Traders can provide liquidity on
Uniswap V3 pools at prices on a log-linear tick space. In particular, consecutive prices are always θ
basis point apart: pi = 1.0001θi , where θ is the tick spacing. For the purpose of the example, we
take θ = 60. Consequently, the current price of 1500.62 corresponds to a tick index of c = 73140.
Figure A.1 illustrates three ticks on grid below and above the current price of ETH/USDT 1500.62.

Figure A.1: ETH/USDT price grid around pc

Price 1473.85 1482.72 1491.64 1500.62 1509.65 1518.73 1527.87

Tick index 72960 73020 73080 73140 73200 73260 73320

Two-sided liquidity provision. Trader A starts out with a capital of USDT 20,000 and wants to
provide liquidity over the price range [1491.64, 1527.87], a range which spans four ticks. Liquidity
provision over a range that includes the current price corresponds to posting quotes on both the bid
and ask side of a traditional limit order book, where the current price of the pool corresponds to
the mid-point of the book.

1. Bid quotes: trader A deposits USDT over the price range [1491.64, 1500.62). This action is
equivalent to submitting a buy limit order with a bid price of 1491.64. An incoming Ether
seller can swap their ETH for the USDT deposited by A, generating price impact until the
limit price of 1491.64 is reached.

2. Ask quotes: at the same time, trader A deposits ETH over three ticks: [1500.62, 1509.65),
[1509.65, 1518.73), and [1518.73, 1527.87). The action corresponds to submitting three sell
limit orders with ask prices 1509.65, 1518.73, and 1527.87, respectively. Incoming Ether buyers
can swap USDT for trader A’s ETH.

In the Uniswap V3 protocol, deposit amounts over each tick [pi , pi+1 ) must satisfy
 
1 1
ETH deposit over [pi , pi+1 ): xi = L √ − √ (A.1)
pi pi+1
√ √
USDT deposit over [pi , pi+1 ): yi = L ( pi+1 − pi ) , (A.2)

37
where L (“liquidity units”) is a scaling factor proportional to the capital committed to the liquidity
position. The scaling factor L is pinned down by setting the total committed capital equal to the
P P
sum of the positions (in USDT), that is pc i xi + i yi . In our example,

1 1
 √ √ 
1500.62 × LA × √ −√ + LA × 1500.62 − 1491.64 = 20000, (A.3)
1500.62 1527.87

leading to LA = 43188.6. We the value of LA into (A.1) and conclude that trader A deposits
5,013.38 USDT over [1491.64, 1500.62) and ETH 9.99 over [1500.62, 1527.87) (approximately ETH
3.33 over each tick size covered).

One-sided liquidity provision. Trader B has USDT 20,000 and wants to post liquidity over the range
[1509.65, 1527.87), which does not include the current price. This action corresponds to posting ask
quotes to sell ETH deep in the book, at price levels 1518.73 and 1527.87. Liquidity is not “active”
– that is, the quotes are not filled – until the existing depth at 1509.65 is consumed by incoming
trades.
We use equation (A.1) to solve for the amount of liquidity units provided by B:
 
1 1
1500.62 × LB × √ −√ = 20000, (A.4)
1509.65 1527.87

which leads to LB = 86589.4. Trader B deposits 6.67 ETH on each of the two ticks covered by the
chosen range.

38
 Figure A.2: ETH/USDT pool state after liquidity provision choices

Figure A.2 illustrates market depth after A and B deposit liquidity in the pool. The current
price of the pool is equivalent to a midpoint in traditional limit order markets. The “ask side” of the
pool is deeper, consistent with both liquidity providers choosing ranges skewed towards prices above
the current midpoint. Liquidity is uniformly provided over ticks – that is, each trader deposits an
equal share of their capital at each price tick covered by their price range.

Trading, fees, and price impact. Suppose now that a trader C wants to buy 10 ETH from the pool.
For each tick interval [pi , pi+1 ), price impact is computed using a constant product function over
virtual reserves:  
L √
x+ √ (y + L pi ) = L2 , (A.5)
pi+1 | {z }
| {z } Virtual USDT reserves
Virtual ETH reserves

where x and y are the actual ETH and USDT deposits in that tick range, respectively. Virtual
reserves are just a mathematical artifact: they extend the physical (real) deposits as if liquidity
would be uniformly distributed over all possible prices on the real line. Working with constant
product functions over real reserves is not feasible: in our example, the product of real reserves is
zero throughout the order book (since only one asset is deposited in each tick range).

39
 Let τ = 1% denote the pool fee that serves as an additional compensation for liquidity providers.
That is, if the buyer pays to pay ∆y USDT to purchase a quantity ∆x ETH, he needs to effectively
pay ∆y (1 + τ ). As per the Uniswap V3 white paper, liquidity fees are not automatically deposited
back into the pool.

1. Tick 1: [1500.62, 1509.65). Trader C first purchases 3.33 ETH at the first available tick above
the current price (equivalent to the “best ask”). To remove the ETH, he needs to deposit ∆y1
USDT, where ∆y1 solves:

√
 
LA 
3.33 − 3.33 + √ 0 + ∆y1 + LA 1500.62 = L2A , (A.6)
1509.65

which leads to ∆y1 = 5026.19 USDT. Trader C pays an average price of 50216.19/3.33=1507.86
USDT for each unit of ETH purchased. Further, he pays a fee of 50.26 USDT to liquidity
provider A (the only liquidity provider at this tick).
The new current price is given by the ratio of virtual reserves,
√
′ ∆y1 + LA 1500.62
p = LA
= 1509.65, (A.7)
3.33 − 3.33 + √1509.65

that is the next price on the tick grid since C exhausts the entire liquidity on [1500.62, 1509.65).

2. Tick 2: [1509.65, 1518.73). Trader C still needs to purchase 6.67 ETH at the next tick level
(where the depth is 10 ETH). The liquidity level at this tick is LA + LB , that is the sum of
liquidity provided by A and B. To remove the 6.67 ETH from the pool, he needs to deposit
∆y2 , where

√
 
LA + LB  
10 − 6.67 + √ 0 + ∆y2 + (LA + LB ) 1509.65 = (LA + LB )2 . (A.8)
1518.73

It follows that trader C purchases 6.67 ETH by depositing ∆y2 = 10089.12 USDT, at an
average price of 1512.61. The pool price is updated as the ratio of virtual reserves:
√
′′ ∆y2 + (LA + LB ) 1509.65
p = = 1515.7. (A.9)
10 − 6.67 + √L1518.73
A +LB

The updated price is in between the two liquidity ticks, since not all depth on this tick level
was exhausted in the trade. Following the swap, liquidity on the tick range [1509.65, 1518.73)
is composed of both assets: that is 10089.12 USDT and 10-6.67=3.33 ETH.
Finally, trader C pays 100.89 USDT as liquidity fees (1% of the trade size), which are
distributed to A and B proportionally to their liquidity share. That is, A receives a fraction

40
 LA
LA +LB of the total fee (33.57 USDT), whereas B receives 67.32 USDT.

Figure A.3 illustrates the impact of the swap. Within tick [1500.62, 1509.65), A sells 3.33 ETH
and buys 5026 USDT. Unlike on limit order books, the execution does not remove liquidity from
the book. Rather, A’s capital is converted from one token to another and remains available to trade.
This feature underscores the passive nature of liquidity supply on decentralized exchanges. Mapping
the concepts to traditional limit order book, this mechanism would imply that every time a market
maker’s sell order is executed at the ask, a buy order would automatically be placed on the bid side
of the market.
The final price of 1517.70 lies within the tick [1509.65, 1518.73), rather than on its boundary.
Trader C only purchases 6.66 ETH out of 10 ETH available within this price interval. The implication
is that liquidity on [1509.65, 1518.73) contains both tokens: 3.33 ETH (the amount that was not
swapped by C) as well as 10089.12 USDT that C deposited in the pool.

41
Figure A.3: Swap execution and price impact

42
 The bottom panel of Figure A.3 shows the price impact of the swap. From equation (6.15) in
the Uniswap V3 white paper, we can solve for the price within tick [pmin , pmax ) with liquidity L,
following the execution of a buy order of size x:

pmin L2
p (x) = √
2 . (A.10)
L − pmin x

As expected, the price impact of a swap decreases in the liquidity available L – each ETH unit
purchased by C has a smaller impact on the price once tick 1509.65 is crossed and the market
becomes deeper.

43
B Notation summary

Variable Subscripts
Subscript Definition
T and N Pertaining to the token and numeraire assets, respectively.
L and H Pertaining to the low- and high-fee pool, respectively.
LP Pertaining to liquidity providers.
LT Pertaining to liquidity traders.
Exogenous Parameters
Parameters Definition
v, ∆ Common and private values of the token.
ℓ, h Liquidity fee on the low- and high-fee pool.
f Liquidity fee on a non-fragmented pool.
qi Token endowment of liquidity provider i.
φ (q, Q) Distribution density of LP endowment q.
Q Maximum token endowment for LPs.
Γ Gas price on the blockchain.
θ Linear trading rate of small LTs.
λ Poisson arrival rate of large LTs.
Endogenous Quantities
Variable Definition
Lk Equilibrium liquidity supply on exchange k ∈ {L, H}.
qt⋆ Token endowment of the LP who is indifferent between pools.
q Lowest token endowment deposited on the market (from break-even condition).
dk Duration of a liquidity cycle on exchange k.
πk Expected liquidity provider profit on exchange k.

44
C Proofs

Lemma 1

Proof. The per-unit liquidity revenue for pool k is expressed as:

√ √

fk (r + 1) 2∆ − r fk + 1 − 2 fk + 1
L (fk ) = . (C.1)
∆
To explore how the liquidity revenue changes with respect to the pool fee, we differentiate L with respect to f :
√


∂L (f ) (r + 1) 2 −2∆ f + 1 + r + 2 + 3f (r + 2)
=− √ . (C.2)
∂f 4∆ f + 1

Starting with f = 0 and given that ∆ > 1 + r (by Assumption 1), the derivative at f = 0 is positive:

∂L(f ) (r + 1)(2∆ − r − 2)
= > 0, (C.3)
∂f f =0 2∆

indicating that liquidity revenue increases with pool fee at this point. The derivative has roots:
p
−6(r + 2)2 + 8∆2 ± 4 4∆4 + 3∆2 (r + 2)2
f1,2 = , (C.4)
9(r + 2)2

where the smallest root f1 is negative and therefore not relevant. We need to show that the largest root f2 is
always positive, defining the threshold f .
For this, consider the numerator of f2 , labeled g(r, ∆):
p
g(r, ∆) = 8∆2 + 4 4∆4 + 3∆2 (r + 2)2 − 6(r + 2)2 .

This function has three roots in r, all of which are negative: r = −2, r = −2(1 + ∆), and r = 2(∆ − 1). Since
these roots are negative, for r ≥ 0, g does not change sign and it is sufficient to examine g(0, ∆):
p
g(0, ∆) = 8∆2 + 4 ∆2 (∆2 + 3) − 6. (C.5)

This is positive for any ∆ ≥ 1, confirming that the largest root f2 is positive and hence, f exists and is
positive. This completes the proof that the liquidity revenue increases with the pool fee until f and decreases
with pool fee for f > f .

Lemma 2

Proof. The cost of adverse selection for pool k after evaluating equation (11) is
√
√

∆− 1 + f (1 + r) ∆2 + ∆ f + 1 (1 + r) + (f + 1)(r − 2)(r + 1) + (f + 1)3/2 r2 (r + 1)
A (fk ) = v .
3∆
(C.6)

45
We aim to demonstrate that A(f ) decreases as f increases. To do this, we calculate the partial derivative of
A(f ) with respect to f :
√

∂A (f ) (r + 1) −2∆ f + 1 + f (r + 2) + r + 2
= √ <0 (C.7)
∂f 2∆ f + 1
√
The derivative is negative if ∆ > 12 1 + f (2 + r). Given that 12 (2 + r) < 1 + r, it follows that A(f ) decreases
√
for any ∆ > (1 + r) 1 + f , consistent with our assumption on ∆.

Proposition 1
L(l)−L(h)
Proof. First, consider the case in which η > L(l)−L(h)+A(l)−A(h) holds. This implies qt < 0, and consequently,
πL − πH < 0 for all q. Under this scenario, liquidity providers universally favor pool H over pool L. They
supply liquidity on pool H if and only if their participation constraint is satisfied, that is if qi > q h .
L(l)−L(h)
Conversely, if η ≤ L(l)−L(h)+A(l)−A(h) , then qt ≥ 0, allowing for a potentially fragmented equilibrium. If
qt ≥ 0, then πL has a steeper slope compared to πH : profit increases more rapidly with liquidity supply in
pool L than in pool H. There are two potential outcomes.

1. Dominance of pool L. If qt < q ℓ < q h , as shown in the left-hand side panel of the diagram below,
then the low-fee pool L captures the entire market share for any qi that yields positive profits. The
condition q ℓ < q h is equivalent to

C (h) (1 − η) L (h) − ηA (h)

qℓ < qh ⇔ > . (C.8)
C (ℓ) (1 − η) L (ℓ) − ηA (ℓ)

2. Fragmented Market Equilibrium. The right-hand side panel depicts the scenario q h < q ℓ < qt . Here,
liquidity providers with qi in the range (q h , qt ] achieve higher positive profits in pool H, while those
with qi > qt obtain larger profits in pool L. The condition q h < q ℓ is equivalent to

C (h) (1 − η) L (h) − ηA (h)

qh ≤ qℓ ⇔ ≤ . (C.9)
C (ℓ) (1 − η) L (ℓ) − ηA (ℓ)

(a) Single pool market (with fee ℓ) (b) Fragmented market

πk πk
πL πL

πH

πH
qt
qℓ qh qh
q Ch qℓ q
Ch qt
Cℓ
Cℓ

46
 It is crucial to note that configurations where q ℓ < qt < q h or q h < qt < q ℓ are not feasible, as they would
lead to a contradiction where profits are simultaneously positive in one pool and negative in the other at the
indifference point qt .

Proposition 2

Proof. We first note that both qt and q h scale linearly with Γ: that is, there exists Qt > Qh > 0 such that
qt = ΓQt and q h = ΓQh where Qt and Qh are not functions of Γ. Next, we compute the partial derivative of
wℓ with respect to Γ
Γ(Qh −Qt )
∂wℓ Γ(Qh − Qt )e λ (ΓQh Qt + λ(Qh + Qt ))
= < 0, (C.10)
∂Γ λ(λ + ΓQh )2
since Qh < Qt and all other terms are positive.

47
D Just-in-time liquidity

Just-in-time (JIT) liquidity is a strategy that leverages the transparency of orders on the public blockchains.
If a liquidity provider observes an incoming large order that has not been processed by miners and it deems
uninformed in the public mempool, it can conveniently re-arrange transactions and propose a sequence of
actions to sandwich this trade as follows:

1. Add a large liquidity deposit at block position k, at the smallest tick around the current pool price.

2. Let the trade at block position k + 1 execute and receive liquidity fees.

3. Remove or burn any residual un-executed liquidity at block position k + 2.

The mint size is optimally very large (i.e., of the order of hundred of millions USD for liquid pairs), such
that the JIT liquidity provider effectively crowds out the existing liquidity supply and collects most fees for
the trade. That is, the strategy is made possible by pro-rata matching on decentralized exchanges because
with time priority, the JIT provider cannot queue-jump existing liquidity providers. Since the JIT liquidity
provider does not want to passively provide capital, it removes any residual deposit immediately after the
trade.
We identify JIT liquidity events by the following algorithm as in Wan and Adams (2022):

1. Search for mints and burns in the same block, liquidity pool, and initiated by the same wallet address.
The mint needs to occur exactly two block positions ahead of the burn (at positions k and k + 2).

2. Classify the mint and the burn as a JIT event if the transaction in between (at position k + 1) is a
trade in the same liquidity pool.

JIT events are rare in our sample, and account for less than 1% of the traded volume on Uniswap v3.
Further, more than half of them occur in a single pair - USDC-WETH, and in low-fee pools. The Uniswap
Labs provides further discussions on the aggregate impact of JIT liquidity provision here. Regarding the
economic effects, JIT liquidity reduces price impact for incoming trades, but dilutes existing liquidity providers
in the pro-rata markets, and can discourage liquidity supply in the long run.

48
E Impermanent loss measure

We build our measure of impermanent loss in line with the definition of token reserves within a price range in
the Uniswap V3 white paper (Adams, Zinsmeister, Salem, Keefer, and Robinson, 2021) and Section 4.1 in
Heimbach, Schertenleib, and Wattenhofer (2022).
Consider a liquidity provider who supplies L units of liquidity into a pool trading a token x for a token y.
The chosen price range is [pℓ , pu ] with pℓ < pu . Further, the current price of the pool is p0 . We are interested
in computing the impermanent loss at a future point in time, when the price updates to p1 .
From Adams, Zinsmeister, Salem, Keefer, and Robinson (2021), the actual amount of tokens x and y
(“real reserves”) deposited on a Uniswap v3 liquidity pool with a price range [pℓ , pu ] to yield liquidity L are
functions of the current pool price p:
   
L × √1pℓ − √1pu if p ≤ pℓ if p ≤ pℓ
0

 
 
√

√
  
x (p) = L × √1p − √1p if pℓ < p ≤ pu and y (p) = L × p − pℓ if pℓ < p ≤ pu (E.1)
u
√ √


 

L× pu − pℓ if p > pu .

0 
if p > pu

From equation (E.1), the value of the liquidity position at t = 1 is therefore

  


 Lp1 × √1pℓ − √1pu if p1 ≤ pℓ
  √ √ 
Vposition = p1 x (p1 ) + y (p1 ) = L × 2 p1 − √pp1 − pℓ if pℓ < p1 ≤ pu (E.2)
u

 L × √p − √ p




u ℓ if p1 > pu .

Conversely, the value of a strategy where the liquidity provider holds the original token quantities and
marks them to market at the updated price is
  
 Lp1 × √1 − √1 if p0 ≤ pℓ

 pℓ pu
√ 
 
Vhold = p1 x (p0 ) + y (p0 ) = L × p√ 1 +p0
p − p
√1 −
p pℓ if pℓ < p0 ≤ pu (E.3)
0 u

 L × √p − √p




u ℓ if p0 > pu .

The impermanent loss is then defined as the excess return from holding the assets versus providing
liquidity on the decentralized exchange:

Vhold − Vposition
ImpermanentLoss = . (E.4)
Vhold

Empirically, we follow Heimbach, Schertenleib, and Wattenhofer (2022) and compute impermanent loss
for “symmetric” positions around the current pool price, that is pℓ = p0 α−1 and pu = p0 α, with α > 1. We
allow for a time lag of one hour between p0 and p1 .

49