Introduction
Empirical studies of many markets report volatility whose paths are rougher than those of Brownian motion [31 ] . Such roughness can arise in theory when each order sets off very long chains of further orders whose influence fades as a power of the elapsed time. In a scaling limit the exponent of this power law sets the roughness of the order flow, which opposing buy and sell flows then carry into the price [22 , 41 ] . Such order flow has been argued to follow from the absence of arbitrage together with the splitting of large orders during execution [22 , 42 ] . These arguments rest on the behavior of traders. However, it is not a priori clear whether market makers that act purely on new information can produce such order flow through the order book on their own.
In this work we provide one mechanism that allows for this. To reflect algorithm-driven strategies, we treat each market maker as a desk that uses a first-order gradient rule to move the distances of its orders from the mid price toward target distances provided by an unknown oracle. As our model has no takers, we let the mid price move toward the side of the book with less volume near it. Each desk thus keeps correcting its orders against a moving reference. We next assume that each desk samples information over its own characteristic horizon [46 ] .
Since at each moment only the desks that have received new information take a step, order placements are asynchronous. Across all desks the rule therefore takes the form of stochastic gradient descent (SGD) with information-driven (mini-batch) noise.
Adapting prior constructions of the continuum limit of SGD [30 ] , we show that the volume of the order book then follows a conservative stochastic partial differential equation (the Dean–Kawasaki equation [21 , 43 ] ) driven by noise that is white in time and colored across the book, with a spatial correlation set by how desks sample information. As desks only redistribute volume, the volume of the order book is conserved irrespective of the nature of this noise. Further, every update of a desk is in turn a source of new information for the other desks; thus the aggregation of this information raises the chance of further updates across the market with a strength set by the shape of the book. Lastly, even though the horizon of each desk leaves a memory that fades exponentially, the distinct horizons cause the market to retain a more persistent memory that appears in the order flow as a power-law decay. Standard scaling arguments then lead to rough volatility when the order flow approaches instability in the asymptotic limit [22 , 41 ] .
We then proceed to test each step of this mechanism, first on simulated desks and then on six months of Binance trades and order books covering about 790 million orders. In simulation the average changes of the order book agree with the theory, with fluctuations that match the information-driven noise. We further confirm that the market shows the predicted memory and that roughness emerges only when desks sample information over distinct horizons. On Binance the order flow is also rough, and its Hurst parameter lies close to that of simulated desks with the same fitted memory. The order books of the perpetual futures there are consistent with how our desks change the book, as orders in these books partly return to their distances after a move of the mid price and move to neighboring distances (§ 7.3 ).
Related literature
Our work broadly straddles the following lines of existing research.
Rough volatility. Rough volatility was first documented in estimates from realized volatility [31 ] . However, other works argue that such estimates could partly reflect estimation error [12 ] . Alternative justifications of roughness that do not rest on these estimates have also been proposed, such as order flow that is close to instability with power-law memory [22 , 41 ] and the absence of arbitrage [FuK21, JR20]. In our work we provide a different route that relies purely on algorithmic market making, with roughness measured from the order count distributions that emerge from it.
Limit order dynamics. Models of the order book itself usually prescribe how orders arrive and leave, either at random [19 SFGK03] or through a latent supply and demand that evolves by reaction and diffusion [3 , 20 ] . We instead choose to rely on optimization dynamics more consistent with modern liquidity provision [2 , 32 , 51 ] . Unlike other agent-based models [15 , 11 , 48 ] , we use a gradient-based rule consistent with modern learning approaches and information-based sampling [46 ] that guides the actions of fast-reacting desks.
Continuum dynamics. Multiple models of order book evolution have been proposed, but most are consistent with the Brownian-like nature of prices [10 MTB14, TLD+11, YST14] . We instead rely on existing continuum approaches for analyzing optimization dynamics [30 , 52 RVE22, SS20] and modify them to be more reflective of the self-exciting and persistent nature of financial order flow.
Outline
We provide a map for the work that follows. First we introduce the desks and their gradient rule in § 2 , after which we derive the conservative equation for the volume of the order book in § 3 . We then proceed to show how fluctuations of the book drive self-exciting order flow with power-law memory (§ 4 ) and connect this order flow to rough volatility (§ 5 ). § 6 and § 7 test the mechanism on simulated desks and on Binance data respectively. Additional details are provided in the Appendix.
Setting: SGD market makers on the order book
Our model has five parts, each of which is needed to turn the gradient rule of a single desk into the dynamics of a whole book. We first describe the order book (§ 2.1 ) and how its mid price moves (§ 2.2 ), then the information that each desk tracks (§ 2.3 ), the gradient rule by which a desk moves its orders (§ 2.4 ) and finally when a desk acts (§ 2.5 ). § 2.6 then puts these parts together in the order in which they occur.
The order book
We consider an idealized order book that consists solely of market makers and evolves in discrete time with steps n = 0 , 1 , … n = 0, 1, \ldots n = 0 , 1 , … . Each order is characterized by its side σ ∈ { − 1 , + 1 } \sigma\in\{-1, +1\} σ ∈ { − 1 , + 1 } , corresponding to the bid and the ask respectively, and by its distance x ≥ 0 x \ge0 x ≥ 0 from a mid price M n M_n M n , measured in units of a price increment h > 0 h > 0 h > 0 . Conventionally the mid price is defined as p n = 1 2 ( A n + B n ) p_n = \frac{1}{2}(A_n + B_n) p n = 2 1 ( A n + B n ) , the average of the best ask A n A_n A n and the best bid B n B_n B n . However, in our model M n M_n M n is a reference price around which all market makers quote. In this setup, if a n a_n a n and b n b_n b n denote the distances of the closest ask and the closest bid from M n M_n M n , then A n = M n + h a n A_n = M_n + ha_n A n = M n + h a n and B n = M n − h b n B_n = M_n - hb_n B n = M n − h b n , so that p n = M n + h 2 ( a n − b n ) p_n = M_n + \frac{h}{2}(a_n - b_n) p n = M n + 2 h ( a n − b n ) , with equality holding if and only if a n = b n a_n = b_n a n = b n .
Our choice follows standard models of market making in which market makers quote around a reference price [2 , 32 ] that in practice is made to account for the imbalance of volume between the two sides around the conventional mid price [Sto18]. Under this setup the mid price moves toward the side with less volume [Sto18], as made precise in (2) . Further, despite being an idealized quantity it is close to the conventional mid price as it differs from the latter by at most half the spread
∣ p n − M n ∣ = h 2 ∣ a n − b n ∣ ≤ 1 2 ( A n − B n ) . \lvert p_n - M_n \rvert= \frac{h}{2}\lvert a_n - b_n\rvert\le\frac{1}{2}(A_n - B_n). ∣ p n − M n ∣ = 2 h ∣ a n − b n ∣ ≤ 2 1 ( A n − B n ) .
The two are thus nearly identical in highly liquid markets with a narrow spread.
In a real order book each market-making participant, which we refer to as a desk d i d_i d i , holds at step n n n a number Q i ( n ) Q_i(n) Q i ( n ) of orders on either side of the book. Each order is a point mass of volume at its distance from the mid price, and desks change the number, sizes, sides and distances of their orders over time.
In our model we represent each participant that quotes on both sides of the book by two desks, one for each side, such that each desk d i d_i d i quotes on a single side σ i \sigma_i σ i . As volume that a participant shifts between its two desks leaves the total volume of the book unchanged, our framework still applies to the book as a whole. We further let each desk hold a fixed number L L L of orders of equal volume m i m_i m i , which discretizes the volume of the desk into equal parts such that a real order corresponds to one or more of these parts at the same distance. Desks in our model only move their orders, which on an exchange corresponds to cancelling an order and immediately placing the same volume at a new distance. Our idealized market thus has no other arrivals, cancellations or annihilations of orders.1 As the volume of a liquid book spreads over many ticks we treat distances as continuous, so that desk d i d_i d i holds its orders o i , 1 , … , o i , L o_{i,1},\ldots,o_{i,L} o i , 1 , … , o i , L at the distances x i ( n ) = ( x i , 1 ( n ) , … , x i , L ( n ) ) ∈ [ 0 , ∞ ) L \boldsymbol{x}_i(n)=(x_{i,1}(n),\ldots,x_{i,L}(n))\in[0,\infty)^L x i ( n ) = ( x i , 1 ( n ) , … , x i , L ( n )) ∈ [ 0 , ∞ ) L and the volume on side σ \sigma σ at step n n n is the measure
V n σ = ∑ i : σ i = σ m i ∑ j = 1 L δ x i , j ( n ) , V_n^\sigma=\sum_{i:\sigma_i=\sigma}m_i\sum_{j=1}^{L}\delta_{x_{i,j}(n)}, V n σ = i : σ i = σ ∑ m i j = 1 ∑ L δ x i , j ( n ) ,
which places the volume m i m_i m i of each order of desk d i d_i d i at its distance. The total volume V n + ( [ 0 , ∞ ) ) + V n − ( [ 0 , ∞ ) ) V_n^+([0,\infty))+V_n^-([0,\infty)) V n + ([ 0 , ∞ )) + V n − ([ 0 , ∞ )) of the book is constant in n n n . In our simulations participants do not shift volume between their two desks, such that the volume on each side is constant as well, which we normalize to one.
How the mid price moves
We first measure how much volume sits near the mid price on each side. We weigh each order by its volume and by its distance from the mid price as
Z n σ = ∑ i : σ i = σ m i ∑ j = 1 L e − x i , j ( n ) / ℓ , (1) Z_n^\sigma=\sum_{i:\sigma_i=\sigma}m_i\sum_{j=1}^{L}e^{-x_{i,j}(n)/\ell}, \tag*{(1)} Z n σ = i : σ i = σ ∑ m i j = 1 ∑ L e − x i , j ( n ) / ℓ , (1)
such that an order at the mid price itself (x = 0 x=0 x = 0 ), which we call the touch, counts with its full volume while the weight of an order decays with its distance over a length ℓ > 0 \ell>0 ℓ > 0 . We chose this soft weighting rather than counting only the volume within a fixed distance, as a fixed cutoff would make Z n σ Z_n^\sigma Z n σ jump whenever an order crosses it. Instead, the chosen soft weighting allows the mid price to respond smoothly to every move. As each side is individually normalized to have total volume one, 0 ≤ Z n σ ≤ 1 0\leq Z_n^\sigma\leq1 0 ≤ Z n σ ≤ 1 , and thus Z n σ Z_n^\sigma Z n σ naturally represents the share of the volume of side σ \sigma σ near the mid price. We relate this share of the volume to the movement of the mid price through
θ n = − c ( Z n + − Z n − ) , M n + 1 = M n + h θ n , (2) \theta_n=-c\left(Z_n^+-Z_n^-\right),\qquad M_{n+1}=M_n+h\theta_n, \tag*{(2)} θ n = − c ( Z n + − Z n − ) , M n + 1 = M n + h θ n , (2)
where c > 0 c>0 c > 0 sets how far the mid price moves for a given imbalance of volume near it. The above linear response is consistent to first order with empirical observations of how moves of the mid price depend on the imbalance of volume [13 , 17 , 27 ] . Our weighting additionally accounts for the contribution of volume at larger distances from the mid price [XGH20].
Since all distances are measured from the mid price, every order shifts with each move of the mid price. Each order of desk d i d_i d i thus moves as
x i , j ← max { x i , j − σ i θ n , 0 } , (3) x_{i,j}\leftarrow\max\{x_{i,j}-\sigma_i\theta_n,0\}, \tag*{(3)} x i , j ← max { x i , j − σ i θ n , 0 } , (3)
such that orders on the side toward which the mid price moves come closer to it while orders on the other side move away. An order that the mid price reaches stays at the touch, as in our idealized market it can neither be executed nor cross to the other side of the book.
Desks and their information
Next, we group the set of all desks into K K K clusters, where cluster k k k contains the desks that sample information over a characteristic horizon T k = T 1 Ω k − 1 T_k=T_1\Omega^{k-1} T k = T 1 Ω k − 1 with a ratio Ω > 1 \Omega>1 Ω > 1 [18 , 46 MDD^{+}97] . As a large part of the liquidity in modern markets is provided by fast desks [51 ] , we let the volume of a cluster decrease as a power of its horizon. The desks of cluster k k k thus together hold the volume
w k = T k − α ∑ l = 1 K T l − α , α > 0 , (4) w_k = \frac{T_k^{-\alpha}}{\sum_{l=1}^{K} T_l^{-\alpha}}, \qquad\alpha> 0, \tag*{(4)} w k = ∑ l = 1 K T l − α T k − α , α > 0 , (4)
on each side, such that the volumes of all clusters sum to one. For simplicity each of the N N N desks of cluster k k k on a given side holds the same volume, such that m i = w k / ( N L ) m_i = w_k/(NL) m i = w k / ( N L ) . In a real market this amounts to fixing the resolution at which we describe the volume of each cluster.
The desks of cluster k k k then track the moving averages of two quantities, which reflect two economic realities.
Market makers adjust their quotes to the recent trend of the mid price, as a predictable short-term drift exposes them to adverse selection [16 ] . The first moving average P k P_k P k therefore follows this trend over the horizon T k T_k T k .
Activity in order books is to a large extent self-exciting, such that recent updates of other participants are themselves a source of information for further updates [25 , 35 ] . The second moving average U k U_k U k therefore follows the recent size of the updates of all desks, which measures market activity.
At every step both are updated as
P k ← P k + θ n − P k T k , U k ← U k + I n − 1 − U k T k , (5) P_k \leftarrow P_k + \frac{\theta_n-P_k}{T_k}, \qquad U_k \leftarrow U_k + \frac{I_{n-1}-U_k}{T_k}, \tag*{(5)} P k ← P k + T k θ n − P k , U k ← U k + T k I n − 1 − U k , (5)
where I n − 1 I_{n-1} I n − 1 measures how strongly the updates of the previous step moved the mid price, as defined in (7) below.
The gradient rule
We now state the exact optimization rule. Each desk d i d_i d i of cluster k k k receives target distances x i ∗ ∈ [ 0 , ∞ ) L x_i^* \in[0,\infty)^L x i ∗ ∈ [ 0 , ∞ ) L for its orders from an unknown oracle, which stands for the desk’s own view of where its orders should sit. As stated earlier, following the recent trend of the mid price, it further leans its orders such that in an upward trend its asks move away from the mid price while its bids move closer. We express both through the loss
L i ( x ) = 1 2 ∑ j = 1 L ( x j − x i , j ∗ ) 2 − σ i λ k ∑ j = 1 L x j , λ k = Π P k γ , (6) \mathcal{L}_i(x) = \frac{1}{2}\sum_{j=1}^{L}(x_j-x_{i,j}^*)^2-\sigma_i\lambda_k\sum_{j=1}^{L}x_j,\qquad\lambda_k=\frac{\Pi P_k}{\gamma}, \tag*{(6)} L i ( x ) = 2 1 j = 1 ∑ L ( x j − x i , j ∗ ) 2 − σ i λ k j = 1 ∑ L x j , λ k = γ Π P k , (6)
where Π ≥ 0 \Pi\ge0 Π ≥ 0 sets how strongly the desk leans with the trend and dividing by γ \gamma γ makes the lean of a single step equal to Π P k \Pi P_k Π P k . When the desk acts, it takes a single gradient step of size γ ∈ ( 0 , 1 ) \gamma\in(0,1) γ ∈ ( 0 , 1 ) as
x i ← max { x i − γ ∇ L i ( x i ) , 0 } = max { x i − γ ( x i − x i ∗ ) + σ i Π P k , 0 } . x_i \leftarrow\max\{x_i-\gamma\nabla\mathcal{L}_i(x_i),0\}=\max\{x_i-\gamma(x_i-x_i^*)+\sigma_i\Pi P_k,0\}. x i ← max { x i − γ ∇ L i ( x i ) , 0 } = max { x i − γ ( x i − x i ∗ ) + σ i Π P k , 0 } .
Notice that as this step affects the volume near the mid price, the next move of the mid price in turn changes by c m i ∣ Δ Z i ∣ c m_i|\Delta Z_i| c m i ∣Δ Z i ∣ , where m i Δ Z i m_i\Delta Z_i m i Δ Z i is the change of Z n n e w Z_n^{\mathrm{new}} Z n new caused by desk d i d_i d i with
Δ Z i = ∑ j ( e − x i , j n e w / ℓ − e − x i , j o l d / ℓ ) . \Delta Z_i=\sum_j\left(e^{-x_{i,j}^{\mathrm{new}}/\ell}-e^{-x_{i,j}^{\mathrm{old}}/\ell}\right). Δ Z i = j ∑ ( e − x i , j new / ℓ − e − x i , j old / ℓ ) .
The updates of all desks acting in step n n n thus change the next move of the mid price by at most
I n = ∑ d i a c t s i n s t e p n c m i ∣ Δ Z i ∣ , (7) I_n=\sum_{d_i\ \mathrm{acts\ in\ step}\ n}c m_i|\Delta Z_i|, \tag*{(7)} I n = d i acts in step n ∑ c m i ∣Δ Z i ∣ , (7)
which we call the size of the updates in step $n.
The mini-batch
Lastly we note that each desk can act only when new information relevant to its oracle arrives. This is modelled by letting each desk act independently with probability
p k = min { p 0 + β U k , 1 } , (8) p_k = \min\{p_0 + \beta U_k, 1\}, \tag*{(8)} p k = min { p 0 + β U k , 1 } , (8)
where k k k denotes its cluster membership, p 0 > 0 p_0 > 0 p 0 > 0 accounts for information that does not come from the desks and β ≥ 0 \beta\ge0 β ≥ 0 sets how strongly desks respond to the recent updates of all desks, including their own. Further, writing ξ i ( n ) ∼ Bernoulli ( p k ) \xi_i(n) \sim\operatorname{Bernoulli}(p_k) ξ i ( n ) ∼ Bernoulli ( p k ) for the indicator that desk d i d_i d i acts in step n n n , its orders move as
x i ← max { x i − ξ i ( n ) γ ∇ L i ( x i ) , 0 } . (9) x_i \leftarrow\max\{x_i - \xi_i(n)\,\gamma\nabla\mathcal{L}_i(x_i), 0\}. \tag*{(9)} x i ← max { x i − ξ i ( n ) γ ∇ L i ( x i ) , 0 } . (9)
These random updates introduce stochasticity into gradient descent akin to mini-batch optimization in machine learning [5 ] . Unlike standard schemes that sample a fixed number of data points per update, our batch consists of the desks that are active and thus grows or shrinks with the size of recent updates.
The market protocol
We initialize P k = U k = I − 1 = 0 P_k = U_k = I_{-1} = 0 P k = U k = I − 1 = 0 and place every order at its target. Though ∇ L i = 0 \nabla\mathcal{L}_i = 0 ∇ L i = 0 for every desk, differing targets for distinct desks create a movement in the mid price and thus kickstart our market dynamics. We now walk through the sequence of events in a single step n n n of the market. First, the mid price moves by θ n \theta_n θ n and shifts all orders following (1) –(3) . Each desk then updates its internal trend P k P_k P k of the mid price and its estimate U k U_k U k of market activity through (5) . Following this update each desk acts independently with probability p k p_k p k in (8) and, if it acts, takes a gradient step toward its targets by (9) . Consequently, all such steps together change the volume near the mid price. This change in turn sets the next move of the mid price as the market passes to step n + 1 n+1 n + 1 , and its size is at most the size of the updates I n I_n I n in (7) . In the following section we study the continuum dynamics of the order book induced by our setting.
The order book in the continuum
We now pass from the discrete protocol (§ 2 ) to a continuum description of the order book. We first write one step of the protocol as an exact sum of a mean and a fluctuation (§ 3.1 ), which leads to a conservative limit as the step size γ \gamma γ tends to zero (§ 3.2 ). In this limit the order book preserves its total volume, such that any redistribution occurs only through continuous flows between depths (§ 3.3 ).
From the protocol to particles
We first place both sides of the book on a single line. For each order o i , j o_{i,j} o i , j let its signed distance be defined as y i , j = σ i x i , j y_{i,j} = \sigma_i x_{i,j} y i , j = σ i x i , j . Clearly bids satisfy y ≤ 0 y \le0 y ≤ 0 and asks y ≥ 0 y \ge0 y ≥ 0 , while the mid price lies exactly at the origin. As long as no order reaches the touch, the shift (3) moves each order as y i , j ← y i , j − θ n y_{i,j} \leftarrow y_{i,j} - \theta_n y i , j ← y i , j − θ n . The movement of the mid price is therefore just a uniform translation of the whole book.
As the drift of each order o i , j o_{i,j} o i , j depends on its index ( i , j ) (i,j) ( i , j ) through its target and cluster, we keep this index in the state of the book. The state at step n n n is then the measure
μ n = ∑ i m i ∑ j = 1 L δ ( y i , j ( n ) , ( i , j ) ) (10) \mu_n = \sum_i m_i \sum_{j=1}^{L} \delta_{(y_{i,j}(n),(i,j))} \tag*{(10)} μ n = i ∑ m i j = 1 ∑ L δ ( y i , j ( n ) , ( i , j )) (10)
together with the moving averages P k P_k P k and U k U_k U k , from which the volume V n ± V_n^\pm V n ± on each side follows by ignoring the indices. Following § 2.6 , the initial state places every order at its target, such that μ 0 = ∑ i m i ∑ j δ ( σ i x i , j ∗ , ( i , j ) ) \mu_0 = \sum_i m_i \sum_j \delta(\sigma_i x_{i,j}^*, (i,j)) μ 0 = ∑ i m i ∑ j δ ( σ i x i , j ∗ , ( i , j )) and P k = U k = I − 1 = 0 P_k = U_k = I_{-1} = 0 P k = U k = I − 1 = 0 .
We next split the step (9) of a desk into its mean and a fluctuation. Recall that desk d i d_i d i acts in step n n n when the Bernoulli variable ξ i ( n ) ∈ { 0 , 1 } \xi_i(n) \in\{0,1\} ξ i ( n ) ∈ { 0 , 1 } equals one, which happens with probability p k p_k p k . Writing
g i , j = − γ σ i ∂ j L i ( x i ) g_{i,j} = -\gamma\sigma_i \partial_j L_i(x_i) g i , j = − γ σ i ∂ j L i ( x i )
for the signed move of order o i , j o_{i,j} o i , j when its desk acts, the expected move of this order given the state F n \mathcal{F}_n F n of the market at step n n n is
E [ ξ i ( n ) g i , j ∣ F n ] = p k g i , j + ( 1 − p k ) ⋅ 0 = p k g i , j . \mathbb{E}[\xi_i(n) g_{i,j} \mid\mathcal{F}_n] = p_k g_{i,j} + (1-p_k)\cdot0 = p_k g_{i,j}. E [ ξ i ( n ) g i , j ∣ F n ] = p k g i , j + ( 1 − p k ) ⋅ 0 = p k g i , j .
The move itself thus splits into this expected value and a fluctuation with mean zero,
ξ i ( n ) g i , j = p k g i , j + ( ξ i ( n ) − p k ) g i , j . \xi_i(n) g_{i,j} = p_k g_{i,j} + (\xi_i(n)-p_k)g_{i,j}. ξ i ( n ) g i , j = p k g i , j + ( ξ i ( n ) − p k ) g i , j .
Notice that the fluctuations of different orders are correlated through the actions of individual desks. We quantify this covariance below.
Lemma 3.1 (Noise of the mini-batch). The fluctuations ζ i , j = ( ξ i ( n ) − p k ) g i , j \zeta_{i,j} = (\xi_i(n)-p_k)g_{i,j} ζ i , j = ( ξ i ( n ) − p k ) g i , j satisfy
E [ ζ i , j ζ i ′ , j ′ ∣ F n ] = 1 { i = i ′ } p k ( 1 − p k ) γ 2 ∂ j L i ( x i ) ∂ j ′ L i ( x i ) . (11) \mathbb{E}[\zeta_{i,j}\zeta_{i',j'} \mid\mathcal{F}_n] = 1\{i=i'\}p_k(1-p_k)\gamma^2\partial_jL_i(x_i)\partial_{j'}L_i(x_i). \tag*{(11)} E [ ζ i , j ζ i ′ , j ′ ∣ F n ] = 1 { i = i ′ } p k ( 1 − p k ) γ 2 ∂ j L i ( x i ) ∂ j ′ L i ( x i ) . (11) Orders of the same desk thus move together, while orders of different desks move independently.
Proof. Given F n \mathcal{F}_n F n , the moves g i , j g_{i,j} g i , j are fixed, as they depend only on the positions, the targets and the trend P k P_k P k at step n n n . Hence
E [ ζ i , j ζ i ′ , j ′ ∣ F n ] = g i , j g i ′ , j ′ E [ ( ξ i ( n ) − p k ) ( ξ i ′ ( n ) − p k ′ ) ∣ F n ] . \mathbb{E}[\zeta_{i,j}\zeta_{i',j'} \mid\mathcal{F}_n] = g_{i,j}g_{i',j'}\mathbb{E}[(\xi_i(n)-p_k)(\xi_{i'}(n)-p_{k'}) \mid\mathcal{F}_n]. E [ ζ i , j ζ i ′ , j ′ ∣ F n ] = g i , j g i ′ , j ′ E [( ξ i ( n ) − p k ) ( ξ i ′ ( n ) − p k ′ ) ∣ F n ] .
For i ≠ i ′ i \ne i' i = i ′ the two indicators are independent given F n \mathcal{F}_n F n , hence the expectation factorizes into E [ ξ i ( n ) − p k ∣ F n ] E [ ξ i ′ ( n ) − p k ′ ∣ F n ] = 0 \mathbb{E}[\xi_i(n)-p_k \mid\mathcal{F}_n]\mathbb{E}[\xi_{i'}(n)-p_{k'} \mid\mathcal{F}_n] = 0 E [ ξ i ( n ) − p k ∣ F n ] E [ ξ i ′ ( n ) − p k ′ ∣ F n ] = 0 . For i = i ′ i=i' i = i ′ it equals the variance E [ ( ξ i ( n ) − p k ) 2 ∣ F n ] = p k ( 1 − p k ) \mathbb{E}[(\xi_i(n)-p_k)^2 \mid\mathcal{F}_n] = p_k(1-p_k) E [( ξ i ( n ) − p k ) 2 ∣ F n ] = p k ( 1 − p k ) of a Bernoulli variable. Lastly, noting that σ i 2 = 1 \sigma_i^2=1 σ i 2 = 1 , we obtain g i , j g i , j ′ = γ 2 ∂ j L i ( x i ) ∂ j ′ L i ( x i ) g_{i,j}g_{i,j'}=\gamma^2\partial_jL_i(x_i)\partial_{j'}L_i(x_i) g i , j g i , j ′ = γ 2 ∂ j L i ( x i ) ∂ j ′ L i ( x i ) . □
We now proceed to study the asymptotic limit of this decomposition into mean and fluctuation terms as the step size γ \gamma γ tends to zero.
The conservative limit
We pass to the continuum in two steps. We first let the step size γ \gamma γ tend to zero and show that the orders, the trends P k P_k P k and the estimates U k U_k U k follow a stochastic differential equation (SDE) with one Brownian motion per desk. Summing this SDE over all orders, we then obtain a conservative stochastic partial differential equation for the volume of the book, of the Dean–Kawasaki type [21 , 43 ] studied in [30 ] .
The time scale and the regime. We consider the time scale on which step n = 0 , 1 , … n=0,1,\ldots n = 0 , 1 , … occurs at the time t = n γ t=n\gamma t = nγ , such that fixing t > 0 t>0 t > 0 and letting γ → 0 + \gamma\to0^+ γ → 0 + allows us to study the continuum limit. We then choose how the parameters of the protocol scale with γ \gamma γ , such that every part of the protocol remains consistent in this limit. If a quantity changes by δ \delta δ in each step, then up to the time t t t it changes by
∑ n < t / γ δ = t γ δ . \sum_{n<t/\gamma}\delta= \frac{t}{\gamma}\delta. n < t / γ ∑ δ = γ t δ .
which stays bounded as γ → 0 + \gamma\to0^{+} γ → 0 + only if δ = O ( γ ) \delta= O(\gamma) δ = O ( γ ) . Following the gradient step (§ 2.4 ), an acting desk moves its orders by − γ ( x i − x i ∗ ) + σ i Π P k -\gamma(x_i-x_i^*)+\sigma_i\Pi P_k − γ ( x i − x i ∗ ) + σ i Π P k , where only the first term is O ( γ ) O(\gamma) O ( γ ) . Further, θ n = − c ( Z n + − Z n − ) \theta_n=-c\left(Z_n^{+}-Z_n^{-}\right) θ n = − c ( Z n + − Z n − ) and its average P k P_k P k are O ( c ) O(c) O ( c ) , while each moving average moves by its gap divided by T k T_k T k . We therefore assume that our regime satisfies the following.
Assumption 3.2 (Regime). There exist constants c ˉ > 0 \bar{c}>0 c ˉ > 0 , τ 1 < ⋯ < τ K \tau_1<\cdots<\tau_K τ 1 < ⋯ < τ K and β ˉ ≥ 0 \bar{\beta}\geq0 β ˉ ≥ 0 , independent of γ \gamma γ , such that
c = γ c ˉ , T k = τ k γ , β = β ˉ γ 2 . c=\gamma\bar{c}, \qquad T_k=\frac{\tau_k}{\gamma}, \qquad\beta=\frac{\bar{\beta}}{\gamma^2}. c = γ c ˉ , T k = γ τ k , β = γ 2 β ˉ .
Notice that the assumption follows from the fact that each scaling is the only one under which its part of the protocol remains consistent in the limit. If c = o ( γ ) c=o(\gamma) c = o ( γ ) the mid price does not move up to the time t t t , whereas if γ = o ( c ) \gamma=o(c) γ = o ( c ) it moves by an unbounded amount. Similarly, T k = τ k / γ T_k=\tau_k/\gamma T k = τ k / γ fixes the horizon of cluster k k k at the time τ k \tau_k τ k . Further, since ∣ e − a / ℓ − e − b / ℓ ∣ ≤ ∣ a − b ∣ / ℓ \left|e^{-a/\ell}-e^{-b/\ell}\right|\leq|a-b|/\ell e − a / ℓ − e − b / ℓ ≤ ∣ a − b ∣/ ℓ for a , b ≥ 0 a,b\geq0 a , b ≥ 0 , an acting desk changes its volume near the mid price by
∣ Δ Z i ∣ ≤ 1 ℓ ∑ j = 1 L ∣ x i , j n e w − x i , j o l d ∣ = O ( γ ) , \left|\Delta Z_i\right|\leq\frac{1}{\ell}\sum_{j=1}^{L}\left|x_{i,j}^{\mathrm{new}}-x_{i,j}^{\mathrm{old}}\right|=O(\gamma), ∣ Δ Z i ∣ ≤ ℓ 1 j = 1 ∑ L x i , j new − x i , j old = O ( γ ) ,
such that I n = O ( γ c ) = O ( γ 2 ) I_n=O(\gamma c)=O(\gamma^2) I n = O ( γ c ) = O ( γ 2 ) . The term β U k \beta U_k β U k thus stays O ( 1 ) O(1) O ( 1 ) without vanishing only if β \beta β is proportional to γ − 2 \gamma^{-2} γ − 2 .
Now, let
θ n = γ v n , P k = γ π k , U k = γ 2 u k , I n = γ 2 ι n , \theta_n=\gamma v_n,\qquad P_k=\gamma\pi_k,\qquad U_k=\gamma^2u_k,\qquad I_n=\gamma^2\iota_n, θ n = γ v n , P k = γ π k , U k = γ 2 u k , I n = γ 2 ι n ,
where v n = − c ˉ ( Z n + − Z n − ) v_n=-\bar{c}\left(Z_n^{+}-Z_n^{-}\right) v n = − c ˉ ( Z n + − Z n − ) is the velocity of the mid price. Then λ k = Π π k \lambda_k=\Pi\pi_k λ k = Π π k and p k = min { p 0 + β ˉ u k , 1 } p_k=\min\{p_0+\bar{\beta}u_k,1\} p k = min { p 0 + β ˉ u k , 1 } do not depend on γ \gamma γ .
Next, we keep the protocol away from the points at which its functions are not smooth, as the limit requires Lipschitz coefficients. The protocol contains three such functions,
u ↦ min { p 0 + β ˉ u , 1 } , x ↦ max { x , 0 } , p ↦ p ( 1 − p ) . u\mapsto\min\{p_0+\bar{\beta}u,1\},\qquad x\mapsto\max\{x,0\},\qquad p\mapsto\sqrt{p(1-p)}. u ↦ min { p 0 + β ˉ u , 1 } , x ↦ max { x , 0 } , p ↦ p ( 1 − p ) .
The first two have a kink, while the standard deviation p ↦ p ( 1 − p ) p\mapsto\sqrt{p(1-p)} p ↦ p ( 1 − p ) of ξ i ( n ) ∼ Bernoulli ( p ) \xi_i(n)\sim\operatorname{Bernoulli}(p) ξ i ( n ) ∼ Bernoulli ( p ) is not Lipschitz continuous at p = 1 p=1 p = 1 .
Assumption 3.3 (Interior). At every step,
(a) p 0 + β ˉ u k ≤ p ˉ p_0+\bar{\beta}u_k\leq\bar{p} p 0 + β ˉ u k ≤ p ˉ for every cluster k k k and some p ˉ < 1 \bar{p}<1 p ˉ < 1 ;
(b) x i , j − σ i θ n > 0 x_{i,j}-\sigma_i\theta_n>0 x i , j − σ i θ n > 0 and x i , j − ξ i ( n ) γ ∂ j L i ( x i ) > 0 x_{i,j}-\xi_i(n)\gamma\partial_jL_i(x_i)>0 x i , j − ξ i ( n ) γ ∂ j L i ( x i ) > 0 for every order o i , j o_{i,j} o i , j .
Under (a) the standard deviation is Lipschitz continuous. Under (b) a move of the mid price shifts every order by − θ n -\theta_n − θ n in signed distances. Both conditions concern the path of the protocol rather than its parameters, and we check them along simulated paths in § 6 .2
The protocol in signed distances. We now write one step of the protocol as an explicit map of the state of the market, such that we can compute its mean and fluctuation. Let y i , j = σ i x i , j y_{i,j}=\sigma_i x_{i,j} y i , j = σ i x i , j , and let k ( i ) k(i) k ( i ) denote the cluster of desk d i d_i d i . Since ∂ j L i ( x i ) = x i , j − x i , j ∗ − σ i Π π k ( i ) \partial_jL_i(x_i)=x_{i,j}-x_{i,j}^{*}-\sigma_i\Pi\pi_{k(i)} ∂ j L i ( x i ) = x i , j − x i , j ∗ − σ i Π π k ( i ) and σ i 2 = 1 \sigma_i^2=1 σ i 2 = 1 , the move g i , j g_{i,j} g i , j (§ 3.1 ) reads
g i , j = − γ σ i ( x i , j − x i , j ∗ − σ i Π π k ( i ) ) = γ b i , j , b i , j = − ( y i , j − y i , j ∗ ) + Π π k ( i ) . g_{i,j}=-\gamma\sigma_i\left(x_{i,j}-x_{i,j}^{*}-\sigma_i\Pi\pi_{k(i)}\right)=\gamma b_{i,j},\qquad b_{i,j}=-\left(y_{i,j}-y_{i,j}^{*}\right)+\Pi\pi_{k(i)}. g i , j = − γ σ i ( x i , j − x i , j ∗ − σ i Π π k ( i ) ) = γ b i , j , b i , j = − ( y i , j − y i , j ∗ ) + Π π k ( i ) .
Writing Z + − Z − Z^{+}-Z^{-} Z + − Z − in signed distances and expanding the exponentials, we similarly obtain
v = − c ˉ ∑ i σ i m i ∑ j = 1 L e − σ i y i , j / ℓ , v=-\bar{c}\sum_i \sigma_i m_i\sum_{j=1}^{L}e^{-\sigma_i y_{i,j}/\ell}, v = − c ˉ i ∑ σ i m i j = 1 ∑ L e − σ i y i , j / ℓ ,
Δ Z i = ∑ j = 1 L ( e − ( x i , j + γ σ i b i , j ) / ℓ − e − x i , j / ℓ ) = γ a i + O ( γ 2 ) , a i = − σ i ℓ ∑ j = 1 L e − σ i y i , j / ℓ b i , j . \Delta Z_i=\sum_{j=1}^{L}\left(e^{-(x_{i,j}+\gamma\sigma_i b_{i,j})/\ell}-e^{-x_{i,j}/\ell}\right)=\gamma a_i+O(\gamma^2),\qquad a_i=-\frac{\sigma_i}{\ell}\sum_{j=1}^{L}e^{-\sigma_i y_{i,j}/\ell}b_{i,j}. Δ Z i = j = 1 ∑ L ( e − ( x i , j + γ σ i b i , j ) / ℓ − e − x i , j / ℓ ) = γ a i + O ( γ 2 ) , a i = − ℓ σ i j = 1 ∑ L e − σ i y i , j / ℓ b i , j .
The protocol of § 2.6 updates U k U_k U k with I n − 1 I_{n-1} I n − 1 . We therefore let step n n n start just before the desks act, which keeps the order of events and updates U k U_k U k with I n I_n I n . The state X n = ( y i , j , π k , u k ) i , j , k X_n=(y_{i,j},\pi_k,u_k)_{i,j,k} X n = ( y i , j , π k , u k ) i , j , k of the market then evolves as
y i , j ( n + 1 ) = y i , j ( n ) + γ ξ i ( n ) b i , j − γ v , y_{i,j}(n+1)=y_{i,j}(n)+\gamma\xi_i(n)b_{i,j}-\gamma v, y i , j ( n + 1 ) = y i , j ( n ) + γ ξ i ( n ) b i , j − γ v ,
π k ( n + 1 ) = π k ( n ) + γ τ k ( v − π k ( n ) ) , \pi_k(n+1)=\pi_k(n)+\frac{\gamma}{\tau_k}\left(v-\pi_k(n)\right), π k ( n + 1 ) = π k ( n ) + τ k γ ( v − π k ( n ) ) ,
u k ( n + 1 ) = u k ( n ) + γ τ k ( ι n − u k ( n ) ) , ι n = c ˉ γ ∑ i ξ i ( n ) m i ∣ Δ Z i ∣ . u_k(n+1)=u_k(n)+\frac{\gamma}{\tau_k}\left(\iota_n-u_k(n)\right),\qquad\iota_n=\frac{\bar{c}}{\gamma}\sum_i\xi_i(n)m_i\lvert\Delta Z_i\rvert. u k ( n + 1 ) = u k ( n ) + τ k γ ( ι n − u k ( n ) ) , ι n = γ c ˉ i ∑ ξ i ( n ) m i ∣ Δ Z i ∣ .
where v v v is evaluated after the gradient steps. Let F n \mathcal{F}_n F n be the σ \sigma σ -algebra generated by ξ ( 0 ) , … , ξ ( n − 1 ) \xi(0),\ldots,\xi(n-1) ξ ( 0 ) , … , ξ ( n − 1 ) . As the initial state is fixed, X n X_n X n is F n \mathcal{F}_n F n -measurable. Given F n \mathcal{F}_n F n , the indicators ξ i ( n ) \xi_i(n) ξ i ( n ) are independent with E [ ξ i ( n ) ∣ F n ] = p k ( i ) \mathbb{E}[\xi_i(n)\mid\mathcal{F}_n]=p_{k(i)} E [ ξ i ( n ) ∣ F n ] = p k ( i ) .
Mean and fluctuation. Following § 3.1 , we next split each step into its conditional mean and a fluctuation with mean zero, and we compute its covariance with Lemma 3.1 . Summed over many steps, the fluctuation becomes Gaussian with independent increments, from which we obtain an SDE. Taking the conditional expectation and using ∣ Δ Z i ∣ = γ ∣ a i ∣ + O ( γ 2 ) \lvert\Delta Z_i\rvert=\gamma\lvert a_i\rvert+O(\gamma^2) ∣ Δ Z i ∣ = γ ∣ a i ∣ + O ( γ 2 ) , we obtain
E [ X n + 1 − X n ∣ F n ] = γ F ( X n ) + O ( γ 2 ) . \mathbb{E}[X_{n+1}-X_n\mid\mathcal{F}_n]=\gamma F(X_n)+O(\gamma^2). E [ X n + 1 − X n ∣ F n ] = γ F ( X n ) + O ( γ 2 ) .
with the drift F F F given by
F y i , j = p k ( i ) b i , j − v , F π k = v − π k τ k , F u k = 1 τ k ( c ˉ ∑ i p k ( i ) m i ∣ a i ∣ − u k ) . F_{y_{i,j}}=p_{k(i)}b_{i,j}-v,\qquad F_{\pi_k}=\frac{v-\pi_k}{\tau_k},\qquad F_{u_k}=\frac{1}{\tau_k}\left(\bar{c}\sum_i p_{k(i)}m_i\lvert a_i\rvert-u_k\right). F y i , j = p k ( i ) b i , j − v , F π k = τ k v − π k , F u k = τ k 1 ( c ˉ i ∑ p k ( i ) m i ∣ a i ∣ − u k ) .
Let s k = p k ( 1 − p k ) s_k=\sqrt{p_k(1-p_k)} s k = p k ( 1 − p k ) , and let S i S_i S i be the vector with the entry s k ( i ) b i , j s_{k(i)}b_{i,j} s k ( i ) b i , j for each order of d i d_i d i , the entry s k ( i ) c ˉ m i ∣ a i ∣ / τ k s_{k(i)}\bar{c}m_i\lvert a_i\rvert/\tau_k s k ( i ) c ˉ m i ∣ a i ∣ / τ k for each u k u_k u k , and zero otherwise. The fluctuation
ζ n + 1 = 1 γ ( X n + 1 − X n − E [ X n + 1 − X n ∣ F n ] ) = ∑ i ξ i ( n ) − p k ( i ) s k ( i ) S i ( X n ) + O ( γ ) \zeta_{n+1}=\frac{1}{\gamma}\left(X_{n+1}-X_n-\mathbb{E}[X_{n+1}-X_n\mid\mathcal{F}_n]\right)=\sum_i\frac{\xi_i(n)-p_{k(i)}}{s_{k(i)}}S_i(X_n)+O(\gamma) ζ n + 1 = γ 1 ( X n + 1 − X n − E [ X n + 1 − X n ∣ F n ] ) = i ∑ s k ( i ) ξ i ( n ) − p k ( i ) S i ( X n ) + O ( γ )
is then linear in the centered indicators and, by construction, has E [ ζ n + 1 ∣ F n ] = 0 \mathbb{E}[\zeta_{n+1}\mid\mathcal{F}_n]=0 E [ ζ n + 1 ∣ F n ] = 0 . For the entries of the orders, we obtain the covariance (11) divided by γ 2 \gamma^2 γ 2 from Lemma 3.1 . The same computation applies to the entries of u k u_k u k , as their fluctuation is driven by the same centered indicators. Together, we obtain
E [ ζ n + 1 ζ n + 1 T ∣ F n ] = ∑ i S i ( X n ) S i ( X n ) T + O ( γ ) . \mathbb{E}\left[\zeta_{n+1}\zeta_{n+1}^{\mathsf{T}}\mid\mathcal{F}_n\right]=\sum_i S_i(X_n)S_i(X_n)^{\mathsf{T}}+O(\gamma). E [ ζ n + 1 ζ n + 1 T ∣ F n ] = i ∑ S i ( X n ) S i ( X n ) T + O ( γ ) .
Since ζ m + 1 \zeta_{m+1} ζ m + 1 is F k \mathcal{F}_k F k -measurable for m < k m<k m < k , we further obtain from the tower property
E [ ζ m + 1 T ζ k + 1 ] = E [ ζ m + 1 T E [ ζ k + 1 ∣ F k ] ] = 0. \mathbb{E}\left[\zeta_{m+1}^{\mathsf{T}}\zeta_{k+1}\right]=\mathbb{E}\left[\zeta_{m+1}^{\mathsf{T}}\mathbb{E}[\zeta_{k+1}\mid\mathcal{F}_k]\right]=0. E [ ζ m + 1 T ζ k + 1 ] = E [ ζ m + 1 T E [ ζ k + 1 ∣ F k ] ] = 0.
Therefore M n = ∑ m < n ζ m + 1 M_n=\sum_{m<n}\zeta_{m+1} M n = ∑ m < n ζ m + 1 is a martingale with orthogonal increments.
For γ ≪ Δ ≪ 1 \gamma\ll\Delta\ll1 γ ≪ Δ ≪ 1 the state changes little over the Δ / γ \Delta/\gamma Δ/ γ steps from t t t to t + Δ t+\Delta t + Δ . By the central limit theorem we then obtain
γ ∑ n = t / γ ( t + Δ ) / γ − 1 ξ i ( n ) − p k ( i ) s k ( i ) S i ( X n ) ≈ γ Δ γ S i ( X t / γ ) N i = γ S i ( X t / γ ) Δ N i , \gamma\sum_{n=t/\gamma}^{(t+\Delta)/\gamma-1} \frac{\xi_i(n)-p_k(i)}{s_k(i)} S_i(X_n) \approx\gamma\sqrt{\frac{\Delta}{\gamma}} S_i(X_{t/\gamma})\mathcal{N}_i = \sqrt{\gamma}S_i(X_{t/\gamma})\sqrt{\Delta}\mathcal{N}_i, γ n = t / γ ∑ ( t + Δ ) / γ − 1 s k ( i ) ξ i ( n ) − p k ( i ) S i ( X n ) ≈ γ γ Δ S i ( X t / γ ) N i = γ S i ( X t / γ ) Δ N i ,
with independent standard Gaussian variables N i \mathcal{N}_i N i . Since the indicators are independent across steps, these increments are independent over disjoint intervals. From the sampling we therefore obtain that they are the increments of independent standard Brownian motions W i W_i W i , which leads us to the SDE
d X = F ( X ) d t + γ ∑ i S i ( X ) d W i . dX = F(X)\,dt+\sqrt{\gamma}\sum_i S_i(X)\,dW_i. d X = F ( X ) d t + γ i ∑ S i ( X ) d W i .
Without the noise, the SDE becomes an ordinary differential equation (ODE) for X ˉ \bar{X} X ˉ . In both cases the mid price follows from M n + 1 = M n + h γ v n M_{n+1}=M_n+h\gamma v_n M n + 1 = M n + hγ v n ,
d X ˉ d t = F ( X ˉ ) , M ( t ) = M ( 0 ) + h ∫ 0 t v ( s ) d s . \frac{d\bar{X}}{dt}=F(\bar{X}), \qquad M(t)=M(0)+h\int_0^t v(s)\,ds. d t d X ˉ = F ( X ˉ ) , M ( t ) = M ( 0 ) + h ∫ 0 t v ( s ) d s .
Distance from the ODE. We now bound the distance of the protocol and of the SDE from X ˉ \bar{X} X ˉ on a finite horizon T T T . We use that F F F is Lipschitz, such that deviations from X ˉ \bar{X} X ˉ grow at most exponentially, while the fluctuations of T / γ T/\gamma T / γ steps add up to a variance proportional to γ \gamma γ . We need Lipschitz coefficients for both facts. We therefore first show that the protocol stays in a bounded set, on which F F F and S i S_i S i are Lipschitz.
Lemma 3.4 (Bounded state). Let Assumptions 3.2 3.2 3.2 and 3.3 3.3 3.3 hold, let γ ≤ τ 1 \gamma\le\tau_1 γ ≤ τ 1 and T > 0 T>0 T > 0 , and let R T = ( 1 + Π ) c ˉ ( T + 1 ) R_T=(1+\Pi)\bar{c}(T+1) R T = ( 1 + Π ) c ˉ ( T + 1 ) . Then for every step n n n with n γ ≤ T n\gamma\le T nγ ≤ T ,
∣ v ∣ ≤ c ˉ , ∣ π k ∣ ≤ c ˉ , ∣ y i , j − y i , j ∗ ∣ ≤ R T , 0 ≤ u k ≤ u ˉ T = 2 c ˉ ( R T + Π c ˉ ) ℓ . |v|\le\bar{c}, \qquad|\pi_k|\le\bar{c}, \qquad|y_{i,j}-y^*_{i,j}|\le R_T, \qquad0\le u_k\le\bar{u}_T=\frac{2\bar{c}(R_T+\Pi\bar{c})}{\ell}. ∣ v ∣ ≤ c ˉ , ∣ π k ∣ ≤ c ˉ , ∣ y i , j − y i , j ∗ ∣ ≤ R T , 0 ≤ u k ≤ u ˉ T = ℓ 2 c ˉ ( R T + Π c ˉ ) .
Proof. Since 0 ≤ Z σ ≤ 1 0\le Z^\sigma\le1 0 ≤ Z σ ≤ 1 and γ ≤ τ k \gamma\le\tau_k γ ≤ τ k , the velocity and the trends satisfy
∣ v ∣ = c ˉ ∣ Z + − Z − ∣ ≤ c ˉ , ∣ π k ( n + 1 ) ∣ ≤ ( 1 − γ τ k ) ∣ π k ( n ) ∣ + γ τ k ∣ v ∣ ≤ c ˉ , |v|=\bar{c}|Z^+-Z^-|\le\bar{c}, \qquad|\pi_k(n+1)|\le\left(1-\frac{\gamma}{\tau_k}\right)|\pi_k(n)|+\frac{\gamma}{\tau_k}|v|\le\bar{c}, ∣ v ∣ = c ˉ ∣ Z + − Z − ∣ ≤ c ˉ , ∣ π k ( n + 1 ) ∣ ≤ ( 1 − τ k γ ) ∣ π k ( n ) ∣ + τ k γ ∣ v ∣ ≤ c ˉ ,
where the second bound follows by induction from π k ( 0 ) = 0 \pi_k(0)=0 π k ( 0 ) = 0 . For the distances, we obtain from the protocol
∣ y i , j ( n + 1 ) − y i , j ∗ ∣ = ∣ ( 1 − γ ξ i ( n ) ) ( y i , j ( n ) − y i , j ∗ ) + γ ( ξ i ( n ) Π π k ( i ) − v ) ∣ ≤ ∣ y i , j ( n ) − y i , j ∗ ∣ + γ ( 1 + Π ) c ˉ . \begin{aligned}
|y_{i,j}(n+1)-y^*_{i,j}|&=|(1-\gamma\xi_i(n))(y_{i,j}(n)-y^*_{i,j})+\gamma(\xi_i(n)\Pi\pi_{k(i)}-v)|\\
&\le|y_{i,j}(n)-y^*_{i,j}|+\gamma(1+\Pi)\bar{c}.
\end{aligned} ∣ y i , j ( n + 1 ) − y i , j ∗ ∣ = ∣ ( 1 − γ ξ i ( n )) ( y i , j ( n ) − y i , j ∗ ) + γ ( ξ i ( n ) Π π k ( i ) − v ) ∣ ≤ ∣ y i , j ( n ) − y i , j ∗ ∣ + γ ( 1 + Π ) c ˉ .
But since § 2.6 places every order at its target and the first move of the mid price is at most γ c ˉ \gamma\bar{c} γ c ˉ , every order starts within γ c ˉ \gamma\bar{c} γ c ˉ of its target. Hence ∣ y i , j ( n ) − y i , j ∗ ∣ ≤ ( n + 1 ) γ ( 1 + Π ) c ˉ ≤ R T |y_{i,j}(n)-y^*_{i,j}|\le(n+1)\gamma(1+\Pi)\bar{c}\le R_T ∣ y i , j ( n ) − y i , j ∗ ∣ ≤ ( n + 1 ) γ ( 1 + Π ) c ˉ ≤ R T . Lastly, using ∣ b i , j ∣ ≤ R T + Π c ˉ |b_{i,j}|\le R_T+\Pi\bar{c} ∣ b i , j ∣ ≤ R T + Π c ˉ , Assumption 3.3 3.3 3.3 (b) and ∑ i L m i = 2 \sum_i Lm_i=2 ∑ i L m i = 2 , we obtain
0 ≤ ι n ≤ c ˉ γ ∑ i m i ∣ Δ Z i ∣ ≤ c ˉ γ ∑ i m i ∑ j = 1 L γ ∣ b i , j ∣ ℓ ≤ 2 c ˉ ( R T + Π c ˉ ) ℓ . 0\le\iota_n\le\frac{\bar{c}}{\gamma}\sum_i m_i|\Delta Z_i|\le\frac{\bar{c}}{\gamma}\sum_i m_i\sum_{j=1}^{L}\frac{\gamma|b_{i,j}|}{\ell}\le\frac{2\bar{c}(R_T+\Pi\bar{c})}{\ell}. 0 ≤ ι n ≤ γ c ˉ i ∑ m i ∣Δ Z i ∣ ≤ γ c ˉ i ∑ m i j = 1 ∑ L ℓ γ ∣ b i , j ∣ ≤ ℓ 2 c ˉ ( R T + Π c ˉ ) .
The bound on u k u_k u k follows by the same induction as for π k \pi_k π k .
Lemma 3.4 bounds the state of the protocol, while the noise can carry the SDE beyond these bounds. We therefore change F F F and S i S_i S i only outside the bounds (Lemma 3.4), which leaves the protocol unaffected. With the changed coefficients, we compare all three processes using Lipschitz estimates. Let K T K_T K T be the box of states with ∣ π k ∣ ≤ c ˉ |\pi_k|\le\bar{c} ∣ π k ∣ ≤ c ˉ , ∣ y i , j − y i , j ∗ ∣ ≤ R T |y_{i,j}-y^*_{i,j}|\le R_T ∣ y i , j − y i , j ∗ ∣ ≤ R T and 0 ≤ u k ≤ min { u ˉ T , ( p ˉ − p 0 ) / β ˉ } 0 \le u_k \le\min\{\bar{u}_T,(\bar{p}-p_0)/\bar{\beta}\} 0 ≤ u k ≤ min { u ˉ T , ( p ˉ − p 0 ) / β ˉ } . By Lemma 3.4 and Assumption 3.3 (a) the protocol stays in K T K_T K T . On K T K_T K T we have p k ∈ [ p 0 , p ˉ ] p_k \in[p_0,\bar{p}] p k ∈ [ p 0 , p ˉ ] , such that F F F and every S i S_i S i are bounded and Lipschitz continuous with a common constant Λ \Lambda Λ . Outside K T K_T K T we evaluate them at the nearest point of the box,
F ~ ( x ) = F ( pr K T ( x ) ) , S ~ i ( x ) = S i ( pr K T ( x ) ) . \widetilde{F}(x)=F(\operatorname{pr}_{K_T}(x)), \qquad\widetilde{S}_i(x)=S_i(\operatorname{pr}_{K_T}(x)). F ( x ) = F ( pr K T ( x )) , S i ( x ) = S i ( pr K T ( x )) .
where pr K T \operatorname{pr}_{K_T} pr K T clips every coordinate to its interval. Clipping is 1 1 1 -Lipschitz and equals the identity on K T K_T K T . The extensions therefore agree with F F F and S i S_i S i on K T K_T K T and are bounded and Λ \Lambda Λ -Lipschitz everywhere. We write Γ \Gamma Γ for a common bound of ∣ F ~ ∣ |\widetilde{F}| ∣ F ∣ , of ∑ i ∣ S ~ i ∣ 2 \sum_i|\widetilde{S}_i|^2 ∑ i ∣ S i ∣ 2 and of ∣ X n + 1 − X n ∣ / γ |X_{n+1}-X_n|/\gamma ∣ X n + 1 − X n ∣/ γ .
Theorem 3.5 (Continuum limit). Let Assumptions 3.2 and 3.3 hold, let γ ≤ min { τ 1 , 1 } \gamma\le\min\{\tau_1,1\} γ ≤ min { τ 1 , 1 } and T > 0 T>0 T > 0 , and let X ˉ \bar{X} X ˉ and X X X solve d X ˉ = F ~ ( X ˉ ) d t d\bar{X}=\widetilde{F}(\bar{X})\,dt d X ˉ = F ( X ˉ ) d t and d X = F ~ ( X ) d t + γ ∑ i S ~ i ( X ) d W i dX=\widetilde{F}(X)\,dt+\sqrt{\gamma}\sum_i\widetilde{S}_i(X)\,dW_i d X = F ( X ) d t + γ ∑ i S i ( X ) d W i , both started from X 0 X_0 X 0 . Then there is a constant C C C , which depends on T T T , c ˉ \bar{c} c ˉ , Π \Pi Π , β ˉ \bar{\beta} β ˉ , ℓ \ell ℓ , p 0 p_0 p 0 , p ˉ \bar{p} p ˉ , τ 1 , … , τ K \tau_1,\ldots,\tau_K τ 1 , … , τ K and the targets but not on γ \gamma γ , such that
(i) E max n γ ≤ T ∣ X n − X ˉ ( n γ ) ∣ 2 ≤ C γ \mathbb{E}\max_{n\gamma\le T}|X_n-\bar{X}(n\gamma)|^2\le C\gamma E max nγ ≤ T ∣ X n − X ˉ ( nγ ) ∣ 2 ≤ C γ ,
(ii) E sup t ≤ T ∣ X ( t ) − X ˉ ( t ) ∣ 2 ≤ C γ \mathbb{E}\sup_{t\le T}|X(t)-\bar{X}(t)|^2\le C\gamma E sup t ≤ T ∣ X ( t ) − X ˉ ( t ) ∣ 2 ≤ C γ .
On every finite horizon the protocol and the SDE thus stay within O ( γ ) O(\sqrt{\gamma}) O ( γ ) of X ˉ \bar{X} X ˉ .
Proof. We begin with (i). By Lemma 3.4 the protocol stays in K T K_T K T , where F ~ = F \widetilde{F}=F F = F . With the fluctuation ζ n + 1 \zeta_{n+1} ζ n + 1 from above, each step reads
X n + 1 = X n + γ F ~ ( X n ) + γ r n + γ ζ n + 1 , ∣ r n ∣ ≤ Γ 1 γ , ∣ ζ n + 1 ∣ ≤ 2 Γ . X_{n+1}=X_n+\gamma\widetilde{F}(X_n)+\gamma r_n+\gamma\zeta_{n+1}, \qquad|r_n|\le\Gamma_1\gamma,\qquad|\zeta_{n+1}|\le2\Gamma. X n + 1 = X n + γ F ( X n ) + γ r n + γ ζ n + 1 , ∣ r n ∣ ≤ Γ 1 γ , ∣ ζ n + 1 ∣ ≤ 2Γ.
where r n r_n r n is the O ( γ 2 ) O(\gamma^2) O ( γ 2 ) term of the conditional mean divided by γ \gamma γ . As ∣ X ˉ ( s ) − X ˉ ( n γ ) ∣ ≤ Γ ( s − n γ ) |\bar{X}(s)-\bar{X}(n\gamma)|\le\Gamma(s-n\gamma) ∣ X ˉ ( s ) − X ˉ ( nγ ) ∣ ≤ Γ ( s − nγ ) , the solution X ˉ \bar{X} X ˉ satisfies
X ˉ ( ( n + 1 ) γ ) = X ˉ ( n γ ) + γ F ~ ( X ˉ ( n γ ) ) + e n , ∣ e n ∣ ≤ ∫ n γ ( n + 1 ) γ Λ Γ ( s − n γ ) d s = Λ Γ γ 2 2 . \bar{X}((n+1)\gamma)=\bar{X}(n\gamma)+\gamma\widetilde{F}(\bar{X}(n\gamma))+e_n,\qquad|e_n|\le\int_{n\gamma}^{(n+1)\gamma}\Lambda\Gamma(s-n\gamma)\,ds=\frac{\Lambda\Gamma\gamma^2}{2}. X ˉ (( n + 1 ) γ ) = X ˉ ( nγ ) + γ F ( X ˉ ( nγ )) + e n , ∣ e n ∣ ≤ ∫ nγ ( n + 1 ) γ ΛΓ ( s − nγ ) d s = 2 ΛΓ γ 2 .
Subtracting and summing from D 0 = 0 D_0=0 D 0 = 0 , we obtain for the deviation D n = X n − X ˉ ( n γ ) D_n=X_n-\bar{X}(n\gamma) D n = X n − X ˉ ( nγ ) with n γ ≤ T n\gamma\le T nγ ≤ T
∣ D n ∣ ≤ γ Λ ∑ m < n ∣ D m ∣ + n ( Γ 1 + Λ Γ 2 ) γ 2 + γ ∣ M n ∣ ≤ γ Λ ∑ m < n ∣ D m ∣ + ( Γ 1 + Λ Γ 2 ) T γ + γ max l ≤ T / γ ∣ M l ∣ . \begin{aligned}
|D_n|\le\gamma\Lambda\sum_{m<n}|D_m|+n\left(\Gamma_1+\frac{\Lambda\Gamma}{2}\right)\gamma^2+\gamma|M_n| \\
&\le\gamma\Lambda\sum_{m<n}|D_m|+\left(\Gamma_1+\frac{\Lambda\Gamma}{2}\right)T\gamma+\gamma\max_{l\le T/\gamma}|M_l|.
\end{aligned} ∣ D n ∣ ≤ γ Λ m < n ∑ ∣ D m ∣ + n ( Γ 1 + 2 ΛΓ ) γ 2 + γ ∣ M n ∣ ≤ γ Λ m < n ∑ ∣ D m ∣ + ( Γ 1 + 2 ΛΓ ) T γ + γ l ≤ T / γ max ∣ M l ∣.
By the discrete Gronwall inequality and ( 1 + γ Λ ) n ≤ e Λ T (1+\gamma\Lambda)^n\le e^{\Lambda T} ( 1 + γ Λ ) n ≤ e Λ T , we obtain
max n γ ≤ T ∣ D n ∣ ≤ e Λ T ( ( Γ 1 + Λ Γ 2 ) T γ + γ max l ≤ T / γ ∣ M l ∣ ) . \max_{n\gamma\le T}|D_n|\le e^{\Lambda T}\left(\left(\Gamma_1+\frac{\Lambda\Gamma}{2}\right)T\gamma+\gamma\max_{l\le T/\gamma}|M_l|\right). nγ ≤ T max ∣ D n ∣ ≤ e Λ T ( ( Γ 1 + 2 ΛΓ ) T γ + γ l ≤ T / γ max ∣ M l ∣ ) .
As M n M_n M n is a martingale with orthogonal increments, we obtain from Doob’s maximal inequality
E max l ≤ T / γ ∣ M l ∣ 2 ≤ 4 E ∣ M ⌊ T / γ ⌋ ∣ 2 = 4 ∑ m < T / γ E ∣ ζ m + 1 ∣ 2 ≤ 16 Γ 2 T γ . \mathbb{E}\max_{l\le T/\gamma}|M_l|^2\le4\mathbb{E}|M_{\lfloor T/\gamma\rfloor}|^2=4\sum_{m<T/\gamma}\mathbb{E}|\zeta_{m+1}|^2\le\frac{16\Gamma^2T}{\gamma}. E l ≤ T / γ max ∣ M l ∣ 2 ≤ 4 E ∣ M ⌊ T / γ ⌋ ∣ 2 = 4 m < T / γ ∑ E ∣ ζ m + 1 ∣ 2 ≤ γ 16 Γ 2 T .
Squaring the previous bound with ( a + b ) 2 ≤ 2 a 2 + 2 b 2 (a+b)^2\le2a^2+2b^2 ( a + b ) 2 ≤ 2 a 2 + 2 b 2 , we obtain
E max n γ ≤ T ∣ D n ∣ 2 ≤ 2 e 2 Λ T ( ( Γ 1 + Λ Γ 2 ) 2 T 2 γ 2 + 16 Γ 2 T γ ) ≤ C γ . \mathbb{E}\max_{n\gamma\le T}|D_n|^2\le2e^{2\Lambda T}\left(\left(\Gamma_1+\frac{\Lambda\Gamma}{2}\right)^2T^2\gamma^2+16\Gamma^2T\gamma\right)\le C\gamma. E nγ ≤ T max ∣ D n ∣ 2 ≤ 2 e 2Λ T ( ( Γ 1 + 2 ΛΓ ) 2 T 2 γ 2 + 16 Γ 2 T γ ) ≤ C γ .
For (ii), the deviation D ( t ) = X ( t ) − X ˉ ( t ) D(t)=X(t)-\bar{X}(t) D ( t ) = X ( t ) − X ˉ ( t ) satisfies
sup s ≤ t ∣ D ( s ) ∣ ≤ Λ ∫ 0 t sup r ≤ s ∣ D ( r ) ∣ d s + γ sup s ≤ T ∣ N s ∣ , N t = ∑ i ∫ 0 t S ~ i ( X ) d W i . \sup_{s\le t}|D(s)|\le\Lambda\int_0^t\sup_{r\le s}|D(r)|\,ds+\sqrt{\gamma}\sup_{s\le T}|N_s|,\qquad N_t=\sum_i\int_0^t\widetilde{S}_i(X)\,dW_i. s ≤ t sup ∣ D ( s ) ∣ ≤ Λ ∫ 0 t r ≤ s sup ∣ D ( r ) ∣ d s + γ s ≤ T sup ∣ N s ∣ , N t = i ∑ ∫ 0 t S i ( X ) d W i .
By Gronwall’s inequality, Doob’s maximal inequality and the Itô isometry, we then obtain
E sup t ≤ T ∣ D ( t ) ∣ 2 ≤ γ e 2 Λ T E sup t ≤ T ∣ N t ∣ 2 ≤ 4 γ e 2 Λ T E ∫ 0 T ∑ i ∣ S ~ i ( X ) ∣ 2 d t ≤ 4 Γ T e 2 Λ T γ . \mathbb{E}\sup_{t\le T}|D(t)|^2 \le\gamma e^{2\Lambda T}\mathbb{E}\sup_{t\le T}|N_t|^2 \le4\gamma e^{2\Lambda T}\mathbb{E}\int_0^T\sum_i|\widetilde{S}_i(X)|^2\,dt \le4\Gamma T e^{2\Lambda T}\gamma. E t ≤ T sup ∣ D ( t ) ∣ 2 ≤ γ e 2Λ T E t ≤ T sup ∣ N t ∣ 2 ≤ 4 γ e 2Λ T E ∫ 0 T i ∑ ∣ S i ( X ) ∣ 2 d t ≤ 4Γ T e 2Λ T γ .
From the orders to the volume. We now sum the SDE over all orders to obtain an equation for the volume of the book. We want this equation to have a unique solution, as the equation then describes the same book as the orders. Intuitively, each order moves by an SDE whose coefficients depend on the whole book. With Itô’s formula we turn the sum of these moves into an equation for the measure of all orders. Conversely, [30 ] shows that every solution of such an equation started from finitely many point masses arises in this way. Our book enters the framework of [30 ] once the indices become a coordinate and the memory is frozen, such that uniqueness reduces to that of our SDE. We first state this framework in full.
Let D ≥ 1 D\ge1 D ≥ 1 , and let Θ \Theta Θ be a measure space with a finite measure ϑ \vartheta ϑ . Let W W W be a cylindrical Wiener process on L 2 ( Θ , ϑ ) L^2(\Theta,\vartheta) L 2 ( Θ , ϑ ) , that is, a family of independent standard Brownian motions W e W^e W e indexed by an orthonormal basis ( e ) (e) ( e ) of L 2 ( Θ , ϑ ) L^2(\Theta,\vartheta) L 2 ( Θ , ϑ ) . Further, let ( F t ) (\mathcal{F}_t) ( F t ) be the filtration generated by W W W , completed by all subsets of events of probability zero and made right-continuous,
F t = ⋂ s > t F s . \mathcal{F}_t=\bigcap_{s>t}\mathcal{F}_s. F t = s > t ⋂ F s .
Here, completion ensures that changing a process on an event of probability zero keeps it adapted. Right-continuity in turn ensures that the information at time t t t includes what is revealed immediately after. Then, the stochastic integral of an adapted process g t ∈ L 2 ( Θ , ϑ ) g_t\in L^2(\Theta,\vartheta) g t ∈ L 2 ( Θ , ϑ ) is
∫ 0 t ∫ Θ g s ( ς ) W ( d ς , d s ) = ∑ e ∫ 0 t ( ∫ Θ g s e d ϑ ) d W s e . \int_0^t\int_\Theta g_s(\varsigma)\,W(d\varsigma,ds)=\sum_e\int_0^t\left(\int_\Theta g_s e\,d\vartheta\right)dW_s^e. ∫ 0 t ∫ Θ g s ( ς ) W ( d ς , d s ) = e ∑ ∫ 0 t ( ∫ Θ g s e d ϑ ) d W s e .
We compare measures by the Wasserstein distance on the set P 2 ( R D ) \mathcal{P}_2(\mathbb{R}^D) P 2 ( R D ) of probability measures with finite second moment,
W 2 ( ν , ν ′ ) = inf { ( E ∣ Y − Y ′ ∣ 2 ) 1 / 2 : Y ∼ ν , Y ′ ∼ ν ′ } . W_2(\nu,\nu')=\inf\left\{(\mathbb{E}|Y-Y'|^2)^{1/2}:Y\sim\nu,\ Y'\sim\nu'\right\}. W 2 ( ν , ν ′ ) = inf { ( E ∣ Y − Y ′ ∣ 2 ) 1/2 : Y ∼ ν , Y ′ ∼ ν ′ } .
Lastly, the coefficients are a velocity V ( t , x , ν ) ∈ R D \mathcal{V}(t,x,\nu)\in\mathbb{R}^D V ( t , x , ν ) ∈ R D and a noise coefficient G ( t , x , ν , ς ) G(t,x,\nu,\varsigma) G ( t , x , ν , ς ) with G ( t , x , ν , ⋅ ) ∈ L 2 ( Θ , ϑ ) D G(t,x,\nu,\cdot)\in L^2(\Theta,\vartheta)^D G ( t , x , ν , ⋅ ) ∈ L 2 ( Θ , ϑ ) D , both possibly random, from which the diffusion matrix follows as
A ( t , x , ν ) = ( ∫ Θ G a G b d ϑ ) a , b ≤ D . A(t,x,\nu)=\left(\int_\Theta G_aG_b\,d\vartheta\right)_{a,b\le D}. A ( t , x , ν ) = ( ∫ Θ G a G b d ϑ ) a , b ≤ D .
The noise of the equation below is thus white in time and has the spatial covariance
∫ Θ G ( x , ς ) G ( x ′ , ς ) T ϑ ( d ς ) \int_\Theta G(x,\varsigma)G(x',\varsigma)^\mathsf{T}\vartheta(d\varsigma) ∫ Θ G ( x , ς ) G ( x ′ , ς ) T ϑ ( d ς )
between two points x x x and x ′ x' x ′ .
Definition 3.7 (Solutions [30 Definitions 2.2 and 2.6] ). (i) A continuous ( F t ) (\mathcal{F}_t) ( F t ) -adapted process ν t \nu_t ν t in P 2 ( R D ) \mathcal{P}_2(\mathbb{R}^D) P 2 ( R D ) is a solution if, for every φ ∈ C c 2 ( R D ) \varphi\in C_c^2(\mathbb{R}^D) φ ∈ C c 2 ( R D ) and t ≥ 0 t \ge0 t ≥ 0 , almost surely
∫ φ d ν t = ∫ φ d ν 0 + ∫ 0 t ∫ ∇ φ ⋅ V ( s , ⋅ , ν s ) d ν s d s + 1 2 ∫ 0 t ∫ D 2 φ : A ( s , ⋅ , ν s ) d ν s d s + ∫ 0 t ∫ Θ ( ∫ ∇ φ ⋅ G ( s , ⋅ , ν s , ς ) d ν s ) W ( d ς , d s ) . \int\varphi\,d\nu_t=\int\varphi\,d\nu_0+\int_0^t\int\nabla\varphi\cdot V(s,\cdot,\nu_s)\,d\nu_s\,ds
+\frac{1}{2}\int_0^t\int D^2\varphi:A(s,\cdot,\nu_s)\,d\nu_s\,ds
+\int_0^t\int_{\Theta}\left(\int\nabla\varphi\cdot G(s,\cdot,\nu_s,\varsigma)\,d\nu_s\right)W(d\varsigma,ds). ∫ φ d ν t = ∫ φ d ν 0 + ∫ 0 t ∫ ∇ φ ⋅ V ( s , ⋅ , ν s ) d ν s d s + 2 1 ∫ 0 t ∫ D 2 φ : A ( s , ⋅ , ν s ) d ν s d s + ∫ 0 t ∫ Θ ( ∫ ∇ φ ⋅ G ( s , ⋅ , ν s , ς ) d ν s ) W ( d ς , d s ) . (ii) A solution started from ν 0 = ∑ l = 1 J α l δ x l \nu_0=\sum_{l=1}^J\alpha_l\delta_{x^l} ν 0 = ∑ l = 1 J α l δ x l is a superposition solution if ν t = ∑ l = 1 J α l δ X l ( t ) \nu_t=\sum_{l=1}^J\alpha_l\delta_{X^l(t)} ν t = ∑ l = 1 J α l δ X l ( t ) , where the points X l X^l X l solve the coupled SDEs
d X l ( t ) = V ( t , X l ( t ) , ν t ) d t + ∫ Θ G ( t , X l ( t ) , ν t , ς ) W ( d ς , d t ) , X l ( 0 ) = x l . dX^l(t)=V(t,X^l(t),\nu_t)\,dt+\int_{\Theta}G(t,X^l(t),\nu_t,\varsigma)\,W(d\varsigma,dt),\qquad X^l(0)=x^l. d X l ( t ) = V ( t , X l ( t ) , ν t ) d t + ∫ Θ G ( t , X l ( t ) , ν t , ς ) W ( d ς , d t ) , X l ( 0 ) = x l .
In a superposition solution each point mass moves by an SDE whose coefficients see the whole measure. Applying Itô’s formula to φ ( X l ) \varphi(X^l) φ ( X l ) and summing over l l l with the weights α l \alpha_l α l , we obtain the identity in (i) [30 Lemma 2.13] . The identity holds for any measure, while the coupled SDEs require point masses. Uniqueness therefore asks whether every solution started from point masses is of this form, which [30 Theorem 2.16] proves under two conditions.
Theorem 3.8 ([30 Theorem 2.16] ). Assume that
( G 1 ) (G1) ( G 1 ) for every t t t , the restrictions of V V V and G G G to [ 0 , t ] [0,t] [ 0 , t ] are B ( [ 0 , t ] ) ⊗ B ( R D ) ⊗ B ( P 2 ) ⊗ F t \mathcal{B}([0,t])\otimes\mathcal{B}(\mathbb{R}^D)\otimes\mathcal{B}(\mathcal{P}_2)\otimes\mathcal{F}_t B ([ 0 , t ]) ⊗ B ( R D ) ⊗ B ( P 2 ) ⊗ F t -measurable and bounded on compact sets;
( G 2 ) (G2) ( G 2 ) for every T > 0 T>0 T > 0 there is L G L_G L G such that almost surely, for all t ≤ T t\le T t ≤ T , x , x ′ ∈ R D x,x'\in\mathbb{R}^D x , x ′ ∈ R D and ν , ν ′ ∈ P 2 ( R D ) \nu,\nu'\in\mathcal{P}_2(\mathbb{R}^D) ν , ν ′ ∈ P 2 ( R D )
∣ V ( t , x , ν ) − V ( t , x ′ , ν ′ ) ∣ + ∥ G ( t , x , ν , ⋅ ) − G ( t , x ′ , ν ′ , ⋅ ) ∥ L 2 ( ϑ ) ≤ L G ( ∣ x − x ′ ∣ + W 2 ( ν , ν ′ ) ) , \lvert V(t,x,\nu)-V(t,x',\nu')\rvert+\lVert G(t,x,\nu,\cdot)-G(t,x',\nu',\cdot)\rVert_{L^2(\vartheta)}
\le L_G\left(\lvert x-x'\rvert+W_2(\nu,\nu')\right), ∣ V ( t , x , ν ) − V ( t , x ′ , ν ′ )∣ + ∥ G ( t , x , ν , ⋅ ) − G ( t , x ′ , ν ′ , ⋅ ) ∥ L 2 ( ϑ ) ≤ L G ( ∣ x − x ′ ∣ + W 2 ( ν , ν ′ ) ) , ∣ V ( t , 0 , δ 0 ) ∣ + ∥ G ( t , 0 , δ 0 , ⋅ ) ∥ L 2 ( ϑ ) ≤ L G . \lvert V(t,0,\delta_0)\rvert+\lVert G(t,0,\delta_0,\cdot)\rVert_{L^2(\vartheta)}\le L_G. ∣ V ( t , 0 , δ 0 )∣ + ∥ G ( t , 0 , δ 0 , ⋅ ) ∥ L 2 ( ϑ ) ≤ L G . Let ν 0 \nu_0 ν 0 consist of J J J point masses. Then there is a unique solution started from ν 0 \nu_0 ν 0 with sup t ≤ T ∫ ∣ x ∣ J 2 + J d ν t < ∞ \sup_{t\le T}\int\lvert x\rvert^{J^2+J}\,d\nu_t<\infty sup t ≤ T ∫ ∣ x ∣ J 2 + J d ν t < ∞ almost surely for every T T T , and it is a superposition solution.
Condition (G1) makes the integrands adapted, such that the stochastic integrals are defined. Condition (G2) in turn makes the coupled SDEs well posed. In [30 ] , these SDEs are solved by freezing the measure, solving, updating the measure and repeating. Condition (G2) makes the errors of this iteration shrink like ( C T ) m / m ! (CT)^m/m! ( C T ) m / m ! . For uniqueness, [30 ] uses the functional
D ( ν ) = ∫ ∏ l ≠ l ′ ∣ z l − z l ′ ∣ ν ( d z 1 ) ⋯ ν ( d z J + 1 ) , D(\nu)=\int\prod_{l\ne l'}\lvert z^l-z^{l'}\rvert\,\nu(dz^1)\cdots\nu(dz^{J+1}), D ( ν ) = ∫ l = l ′ ∏ ∣ z l − z l ′ ∣ ν ( d z 1 ) ⋯ ν ( d z J + 1 ) ,
which is the expected product of all distances between J + 1 J+1 J + 1 points drawn independently from ν \nu ν . If ν \nu ν has at most J J J point masses, two of the J + 1 J+1 J + 1 points coincide and D ( ν ) = 0 D(\nu)=0 D ( ν ) = 0 , while otherwise D ( ν ) > 0 D(\nu)>0 D ( ν ) > 0 . With an Itô formula for functions of finitely many integrals ∫ φ l d ν \int\varphi_l\,d\nu ∫ φ l d ν , (G2) and Gronwall’s inequality, [30 ] obtains
E D ( ν t ) ≤ C ∫ 0 t E D ( ν s ) d s , D ( ν 0 ) = 0 , \mathbb{E}\,\mathcal{D}(\nu_t) \le C \int_0^t \mathbb{E}\,\mathcal{D}(\nu_s)\,ds,\qquad\mathcal{D}(\nu_0)=0, E D ( ν t ) ≤ C ∫ 0 t E D ( ν s ) d s , D ( ν 0 ) = 0 ,
such that D ( ν t ) = 0 \mathcal{D}(\nu_t)=0 D ( ν t ) = 0 and every solution keeps at most J J J point masses [30 Proposition 2.18 and Lemma 2.21] . Here, the moment bound of order J 2 + J J^2+J J 2 + J makes D ( ν t ) \mathcal{D}(\nu_t) D ( ν t ) integrable. Lastly, the proof of [30 Theorem 2.16] shows that each of these point masses moves by the coupled SDEs, whose solution is unique by (G2). The solution is therefore the superposition solution.
The conditions for our book. We now place our book into this framework and check its two conditions. The indices become a second coordinate, Θ \Theta Θ becomes the set of desks, and the memory enters as a frozen input. Let J J J be the number of orders. We place each index ( i , j ) (i,j) ( i , j ) at its own point q i , j ∈ R q_{i,j}\in\mathbb{R} q i , j ∈ R , and we write σ ( q ) \sigma(q) σ ( q ) , y ∗ ( q ) y^*(q) y ∗ ( q ) , k ( q ) k(q) k ( q ) and 1 i ( q ) \mathbf{1}_i(q) 1 i ( q ) for the side, signed target, cluster and desk indicator of the order at q q q . Between the points q i , j q_{i,j} q i , j we interpolate these functions linearly, which are extended outside this range as constants. The book then becomes the probability measure
ν t = 1 2 ∑ i m i ∑ j = 1 L δ ( y i , j ( t ) , q i , j ) on R 2 , \nu_t=\frac{1}{2}\sum_i m_i\sum_{j=1}^{L}\delta_{(y_{i,j}(t),q_{i,j})}\quad\text{on }\mathbb{R}^2, ν t = 2 1 i ∑ m i j = 1 ∑ L δ ( y i , j ( t ) , q i , j ) on R 2 ,
whose second coordinate never moves. Further, let Θ \Theta Θ be the set of desks with counting measure, such that W ( { d i } , t ) = W i ( t ) W(\{d_i\},t)=W_i(t) W ({ d i } , t ) = W i ( t ) . Let y ˉ = pr ( y , q ) \bar y=\operatorname{pr}(y,q) y ˉ = pr ( y , q ) clip y y y to [ y ∗ ( q ) − R T , y ∗ ( q ) + R T ] [y^*(q)-R_T,y^*(q)+R_T] [ y ∗ ( q ) − R T , y ∗ ( q ) + R T ] , and let χ ˉ = ( π ˉ , u ˉ ) \bar\chi=(\bar\pi,\bar u) χ ˉ = ( π ˉ , u ˉ ) clip a memory value χ = ( π , u ) \chi=(\pi,u) χ = ( π , u ) to the box K T K_T K T . In terms of ν \nu ν , the quantities b b b , v v v and a i a_i a i then read
b ( y , q , χ ) = − ( y ˉ − y ∗ ( q ) ) + Π π ˉ k ( q ) , b(y,q,\chi)=-(\bar y-y^*(q))+\Pi\bar\pi_{k(q)}, b ( y , q , χ ) = − ( y ˉ − y ∗ ( q )) + Π π ˉ k ( q ) ,
v ( ν ) = − 2 c ˉ ∫ σ ( q ) e − σ ( q ) y ˉ / ℓ ν ( d y , d q ) , v(\nu)=-2\bar c\int\sigma(q)e^{-\sigma(q)\bar y/\ell}\,\nu(dy,dq), v ( ν ) = − 2 c ˉ ∫ σ ( q ) e − σ ( q ) y ˉ / ℓ ν ( d y , d q ) ,
a i ( ν , χ ) = − 2 σ i ℓ m i ∫ 1 i ( q ) e − σ ( q ) y ˉ / ℓ b ( y , q , χ ) ν ( d y , d q ) . a_i(\nu,\chi)=-\frac{2\sigma_i}{\ell m_i}\int\mathbf{1}_i(q)e^{-\sigma(q)\bar y/\ell}b(y,q,\chi)\,\nu(dy,dq). a i ( ν , χ ) = − ℓ m i 2 σ i ∫ 1 i ( q ) e − σ ( q ) y ˉ / ℓ b ( y , q , χ ) ν ( d y , d q ) .
For a memory path χ ( t ) = ( π ( t ) , u ( t ) ) \chi(t)=(\pi(t),u(t)) χ ( t ) = ( π ( t ) , u ( t )) we define the coefficients with frozen memory
V χ ( t , ( y , q ) , ν ) = ( p k ( q ) ( u ˉ ( t ) ) b ( y , q , χ ( t ) ) − v ( ν ) 0 ) , \mathcal{V}^{\chi}(t,(y,q),\nu)=
\begin{pmatrix}
p_{k(q)}(\bar u(t))b(y,q,\chi(t))-v(\nu)\\
0
\end{pmatrix}, V χ ( t , ( y , q ) , ν ) = ( p k ( q ) ( u ˉ ( t )) b ( y , q , χ ( t )) − v ( ν ) 0 ) ,
G χ ( t , ( y , q ) , ν , d i ) = ( γ 1 i ( q ) s k ( q ) ( u ˉ ( t ) ) b ( y , q , χ ( t ) ) 0 ) , G^{\chi}(t,(y,q),\nu,d_i)=
\begin{pmatrix}
\sqrt{\gamma}\,\mathbf{1}_i(q)s_{k(q)}(\bar u(t))b(y,q,\chi(t))\\
0
\end{pmatrix}, G χ ( t , ( y , q ) , ν , d i ) = ( γ 1 i ( q ) s k ( q ) ( u ˉ ( t )) b ( y , q , χ ( t )) 0 ) ,
where p k ( u ) = p 0 + β ˉ u k p_k(u)=p_0+\bar\beta u_k p k ( u ) = p 0 + β ˉ u k and s k ( u ) = p k ( u ) ( 1 − p k ( u ) ) s_k(u)=\sqrt{p_k(u)(1-p_k(u))} s k ( u ) = p k ( u ) ( 1 − p k ( u )) . At the point masses of ν \nu ν , these coefficients are the order components of F ~ \widetilde F F and γ S ~ i \sqrt{\gamma}\,\widetilde S_i γ S i .
Lemma 3.9 (Frozen memory). Let χ \chi χ be a continuous ( F t ) (\mathcal{F}_t) ( F t ) -adapted process in R 2 K \mathbb{R}^{2K} R 2 K . Then V χ \mathcal{V}^{\chi} V χ and G χ G^{\chi} G χ satisfy (G1) and (G2) with a constant L G L_G L G that does not depend on χ \chi χ .
Proof. For (G1), the coefficients have the form V χ ( t , z , ν , ω ) = Ψ ( z , ν , χ ( t , ω ) ) \mathcal{V}^{\chi}(t,z,\nu,\omega)=\Psi(z,\nu,\chi(t,\omega)) V χ ( t , z , ν , ω ) = Ψ ( z , ν , χ ( t , ω )) with a continuous function Ψ \Psi Ψ , and similarly for G χ G^{\chi} G χ . A continuous adapted process is progressively measurable. Hence for every t t t the map
( s , z , ν , ω ) ↦ ( z , ν , χ ( s , ω ) ) ↦ Ψ ( z , ν , χ ( s , ω ) ) , s ≤ t , (s,z,\nu,\omega)\mapsto(z,\nu,\chi(s,\omega))\mapsto\Psi(z,\nu,\chi(s,\omega)),\qquad s\le t, ( s , z , ν , ω ) ↦ ( z , ν , χ ( s , ω )) ↦ Ψ ( z , ν , χ ( s , ω )) , s ≤ t ,
is B ( [ 0 , t ] ) ⊗ B ( R 2 ) ⊗ B ( P 2 ) ⊗ F t \mathcal{B}([0,t])\otimes\mathcal{B}(\mathbb{R}^2)\otimes\mathcal{B}(\mathcal{P}_2)\otimes\mathcal{F}_t B ([ 0 , t ]) ⊗ B ( R 2 ) ⊗ B ( P 2 ) ⊗ F t -measurable. Moreover, the coefficients are bounded, as all their factors are evaluated at clipped arguments.
For (G2) we use that a product of bounded Lipschitz functions is Lipschitz,
∣ f g ( z ) − f g ( z ′ ) ∣ ≤ ( ∥ f ∥ ∞ Lip ( g ) + ∥ g ∥ ∞ Lip ( f ) ) ∣ z − z ′ ∣ . |fg(z)-fg(z')| \leq(\|f\|_\infty\operatorname{Lip}(g)+\|g\|_\infty\operatorname{Lip}(f))|z-z'|. ∣ f g ( z ) − f g ( z ′ ) ∣ ≤ ( ∥ f ∥ ∞ Lip ( g ) + ∥ g ∥ ∞ Lip ( f )) ∣ z − z ′ ∣.
The factors σ ( q ) \sigma(q) σ ( q ) , y ∗ ( q ) y^*(q) y ∗ ( q ) , 1 i ( q ) 1_i(q) 1 i ( q ) and p k ( q ) ( u ˉ ) p_{k(q)}(\bar{u}) p k ( q ) ( u ˉ ) are bounded and Lipschitz in q q q by the linear interpolation, with constants set by the smallest gap between the points q i , j q_{i,j} q i , j . Similarly, the factor y ˉ \bar{y} y ˉ is 1-Lipschitz in y y y and bounded, such that b b b is bounded and Lipschitz in ( y , q ) (y,q) ( y , q ) . Further, s k ( q ) ( u ˉ ) s_{k(q)}(\bar{u}) s k ( q ) ( u ˉ ) is bounded and Lipschitz as p k ( q ) ( u ˉ ) ∈ [ p 0 , p ˉ ] p_{k(q)}(\bar{u}) \in[p_0,\bar{p}] p k ( q ) ( u ˉ ) ∈ [ p 0 , p ˉ ] . Hence
∣ V χ ( t , z , ν ) − V χ ( t , z ′ , ν ) ∣ + ∥ G χ ( t , z , ν , ⋅ ) − G χ ( t , z ′ , ν , ⋅ ) ∥ L 2 ( ϑ ) ≤ L 1 ∣ z − z ′ ∣ . |V^\chi(t,z,\nu)-V^\chi(t,z',\nu)|+\|G^\chi(t,z,\nu,\cdot)-G^\chi(t,z',\nu,\cdot)\|_{L^2(\vartheta)}\leq L_1|z-z'|. ∣ V χ ( t , z , ν ) − V χ ( t , z ′ , ν ) ∣ + ∥ G χ ( t , z , ν , ⋅ ) − G χ ( t , z ′ , ν , ⋅ ) ∥ L 2 ( ϑ ) ≤ L 1 ∣ z − z ′ ∣.
Lastly, the coefficients depend on the measure only through v ( ν ) v(\nu) v ( ν ) , whose integrand f ( y , q ) = σ ( q ) e − σ ( q ) y ˉ / ℓ f(y,q)=\sigma(q)e^{-\sigma(q)\bar{y}/\ell} f ( y , q ) = σ ( q ) e − σ ( q ) y ˉ / ℓ is bounded and Lipschitz by the same argument. For any Y ∼ ν Y\sim\nu Y ∼ ν and Y ′ ∼ ν ′ Y'\sim\nu' Y ′ ∼ ν ′ , we obtain from the Cauchy–Schwarz inequality
∣ v ( ν ) − v ( ν ′ ) ∣ = 2 c ˉ ∣ E f ( Y ) − E f ( Y ′ ) ∣ ≤ 2 c ˉ Lip ( f ) E ∣ Y − Y ′ ∣ ≤ 2 c ˉ Lip ( f ) ( E ∣ Y − Y ′ ∣ 2 ) 1 / 2 . |v(\nu)-v(\nu')|=2\bar{c}|\mathbb{E}f(Y)-\mathbb{E}f(Y')|\leq2\bar{c}\operatorname{Lip}(f)\mathbb{E}|Y-Y'|\leq2\bar{c}\operatorname{Lip}(f)(\mathbb{E}|Y-Y'|^2)^{1/2}. ∣ v ( ν ) − v ( ν ′ ) ∣ = 2 c ˉ ∣ E f ( Y ) − E f ( Y ′ ) ∣ ≤ 2 c ˉ Lip ( f ) E ∣ Y − Y ′ ∣ ≤ 2 c ˉ Lip ( f ) ( E ∣ Y − Y ′ ∣ 2 ) 1/2 .
Taking the infimum over ( Y , Y ′ ) (Y,Y') ( Y , Y ′ ) , we obtain ∣ v ( ν ) − v ( ν ′ ) ∣ ≤ 2 c ˉ Lip ( f ) W 2 ( ν , ν ′ ) |v(\nu)-v(\nu')|\leq2\bar{c}\operatorname{Lip}(f)W_2(\nu,\nu') ∣ v ( ν ) − v ( ν ′ ) ∣ ≤ 2 c ˉ Lip ( f ) W 2 ( ν , ν ′ ) . With L G = L 1 + 2 c ˉ Lip ( f ) + sup ∣ V χ ∣ + sup ∥ G χ ∥ L_G=L_1+2\bar{c}\operatorname{Lip}(f)+\sup|V^\chi|+\sup\|G^\chi\| L G = L 1 + 2 c ˉ Lip ( f ) + sup ∣ V χ ∣ + sup ∥ G χ ∥ , both inequalities of (G2) hold. None of these constants involves χ \chi χ , since χ \chi χ enters only through clipped values. □
The book itself generates its memory. We now build this property into the notion of solution and call a pair ( ν , χ ) (\nu,\chi) ( ν , χ ) a coupled solution if ν \nu ν solves Definition 3.7(i) with the coefficients ( V χ , G χ ) (V^\chi,G^\chi) ( V χ , G χ ) and the moment bound (Theorem 3.8), and if χ = ( π , u ) \chi=(\pi,u) χ = ( π , u ) solves
d π k = v ( ν t ) − π k τ k d t , d\pi_k=\frac{v(\nu_t)-\pi_k}{\tau_k}\,dt, d π k = τ k v ( ν t ) − π k d t ,
d u k = 1 τ k ( c ˉ ∑ i p k ( i ) ( u ˉ ) m i ∣ a i ( ν t , χ ) ∣ − u k ) d t + γ τ k ∑ i s k ( i ) ( u ˉ ) c ˉ m i ∣ a i ( ν t , χ ) ∣ d W i , du_k=\frac{1}{\tau_k}\left(\bar{c}\sum_i p_{k(i)}(\bar{u})m_i|a_i(\nu_t,\chi)|-u_k\right)dt+\frac{\sqrt{\gamma}}{\tau_k}\sum_i s_{k(i)}(\bar{u})\bar{c}m_i|a_i(\nu_t,\chi)|\,dW_i, d u k = τ k 1 ( c ˉ i ∑ p k ( i ) ( u ˉ ) m i ∣ a i ( ν t , χ ) ∣ − u k ) d t + τ k γ i ∑ s k ( i ) ( u ˉ ) c ˉ m i ∣ a i ( ν t , χ ) ∣ d W i ,
from π k ( 0 ) \pi_k(0) π k ( 0 ) and u k ( 0 ) u_k(0) u k ( 0 ) of the initial state.
Proposition 3.10 (A variant of the Dean–Kawasaki equation). Let X X X solve the SDE with the extended coefficients, and let ν t \nu_t ν t be the measure of its orders. Then ( ν , ( π , u ) ) (\nu,(\pi,u)) ( ν , ( π , u )) is a coupled solution, and for every φ ∈ C b 2 ( R ) \varphi\in C_b^2(\mathbb{R}) φ ∈ C b 2 ( R )
d ∑ i m i ∑ j = 1 L φ ( y i , j ) = ∑ i m i ∑ j = 1 L ( φ ′ ( y i , j ) V i , j + γ 2 φ ′ ′ ( y i , j ) G i , j 2 ) d t + γ ∑ i ( m i ∑ j = 1 L φ ′ ( y i , j ) G i , j ) d W i (12) d\sum_i m_i\sum_{j=1}^{L}\varphi(y_{i,j})=\sum_i m_i\sum_{j=1}^{L}\left(\varphi'(y_{i,j})V_{i,j}+\frac{\gamma}{2}\varphi''(y_{i,j})G_{i,j}^2\right)dt+\sqrt{\gamma}\sum_i\left(m_i\sum_{j=1}^{L}\varphi'(y_{i,j})G_{i,j}\right)dW_i \tag*{(12)} d i ∑ m i j = 1 ∑ L φ ( y i , j ) = i ∑ m i j = 1 ∑ L ( φ ′ ( y i , j ) V i , j + 2 γ φ ′′ ( y i , j ) G i , j 2 ) d t + γ i ∑ ( m i j = 1 ∑ L φ ′ ( y i , j ) G i , j ) d W i (12) with V i , j = p k ( i ) b i , j − v V_{i,j}=p_{k(i)}b_{i,j}-v V i , j = p k ( i ) b i , j − v and G i , j = s k ( i ) b i , j G_{i,j}=s_{k(i)}b_{i,j} G i , j = s k ( i ) b i , j at the clipped state. In particular, the total volume ∑ i L m i = 2 \sum_i Lm_i=2 ∑ i L m i = 2 is constant. Every coupled solution equals ( ν , ( π , u ) ) (\nu,(\pi,u)) ( ν , ( π , u )) .
Integrating by parts in y y y , we obtain for the volume μ t = 2 ν t \mu_t=2\nu_t μ t = 2 ν t a variant of the Dean–Kawasaki equation [21 , 43 ]
d μ t = − ∂ y ( ν μ t ) d t + γ 2 ∂ y 2 ( G 2 μ t ) d t − γ ∑ i ∂ y ( G 1 i μ t ) d W i . d\mu_t=-\partial_y(\nu\mu_t)\,dt+\frac{\gamma}{2}\partial_y^2(G^2\mu_t)\,dt-\sqrt{\gamma}\sum_i\partial_y(G1_i\mu_t)\,dW_i. d μ t = − ∂ y ( ν μ t ) d t + 2 γ ∂ y 2 ( G 2 μ t ) d t − γ i ∑ ∂ y ( G 1 i μ t ) d W i .
Every term on its right side is a derivative in y y y , such that the equation is conservative.
Proof. By the SDE, each order moves as d y i , j = V i , j d t + γ G i , j d W i dy_{i,j} = \mathcal{V}_{i,j}\,dt + \sqrt{\gamma}\,G_{i,j}\,dW_i d y i , j = V i , j d t + γ G i , j d W i , and we obtain from Itô’s formula
d φ ( y i , j ) = ( φ ′ ( y i , j ) V i , j + γ 2 φ ′ ′ ( y i , j ) G i , j 2 ) d t + γ φ ′ ( y i , j ) G i , j d W i . d\varphi(y_{i,j}) = \left(\varphi'(y_{i,j})\mathcal{V}_{i,j} + \frac{\gamma}{2}\varphi''(y_{i,j})G_{i,j}^{2}\right)dt + \sqrt{\gamma}\,\varphi'(y_{i,j})G_{i,j}\,dW_i. d φ ( y i , j ) = ( φ ′ ( y i , j ) V i , j + 2 γ φ ′′ ( y i , j ) G i , j 2 ) d t + γ φ ′ ( y i , j ) G i , j d W i .
Multiplying by m i m_i m i and summing over all orders, we obtain (12) . Choosing φ ≡ 1 \varphi\equiv1 φ ≡ 1 , we see that every term on the right vanishes, so the total volume is constant. Since V i , j \mathcal{V}_{i,j} V i , j and γ G i , j \sqrt{\gamma}G_{i,j} γ G i , j equal V χ \mathcal{V}^{\chi} V χ and G χ ( ⋅ , d i ) G^{\chi}(\cdot,d_i) G χ ( ⋅ , d i ) at the point ( y i , j , q i , j ) (y_{i,j},q_{i,j}) ( y i , j , q i , j ) , the same computation for φ ∈ C c 2 ( R 2 ) \varphi\in C_c^2(\mathbb{R}^2) φ ∈ C c 2 ( R 2 ) gives the identity (Definition 3.7(i)). Further, the moment bound holds because the coefficients are bounded, while the memory equations are the π \pi π and u u u rows of the SDE. Hence ( ν , ( π , u ) ) (\nu,(\pi,u)) ( ν , ( π , u )) is a coupled solution.
For uniqueness, let ( ν ′ , χ ′ ) (\nu',\chi') ( ν ′ , χ ′ ) be a coupled solution. Freezing χ ′ \chi' χ ′ , we obtain from Lemma 3.9 and Theorem 3.8 that ν ′ \nu' ν ′ is the superposition solution for the coefficients ( V χ ′ , G χ ′ ) (\mathcal{V}^{\chi'},G^{\chi'}) ( V χ ′ , G χ ′ ) ,
ν t ′ = 1 2 ∑ i m i ∑ j = 1 L δ ( y i , j ′ ( t ) , q i , j ) , d y i , j ′ = V χ ′ ( t , ( y i , j ′ , q i , j ) , ν t ′ ) d t + G χ ′ ( t , ( y i , j ′ , q i , j ) , ν t ′ , d i ) d W i , \nu'_t = \frac{1}{2}\sum_i m_i\sum_{j=1}^{L}\delta_{(y'_{i,j}(t),q_{i,j})},\qquad
dy'_{i,j} = \mathcal{V}^{\chi'}(t,(y'_{i,j},q_{i,j}),\nu'_t)\,dt + G^{\chi'}(t,(y'_{i,j},q_{i,j}),\nu'_t,d_i)\,dW_i, ν t ′ = 2 1 i ∑ m i j = 1 ∑ L δ ( y i , j ′ ( t ) , q i , j ) , d y i , j ′ = V χ ′ ( t , ( y i , j ′ , q i , j ) , ν t ′ ) d t + G χ ′ ( t , ( y i , j ′ , q i , j ) , ν t ′ , d i ) d W i ,
where the second coordinate stays at q i , j q_{i,j} q i , j as its coefficients vanish. The state X ′ = ( y ′ , π ′ , u ′ ) X'=(y',\pi',u') X ′ = ( y ′ , π ′ , u ′ ) therefore solves the order rows of the SDE, while χ ′ \chi' χ ′ solves its memory rows. Hence X ′ X' X ′ solves the SDE with the extended coefficients. These coefficients are Lipschitz, such that the SDE has a unique strong solution. Consequently X ′ = X X'=X X ′ = X almost surely and ( ν ′ , χ ′ ) = ( ν , ( π , u ) ) (\nu',\chi')=(\nu,(\pi,u)) ( ν ′ , χ ′ ) = ( ν , ( π , u )) . □
We next follow how the conservative equation moves volume between distances.
Conservation
In the queueing models of the order book [1 , 3 , 14 , 19 GPW+13, 38 SFGK03] , the mid price moves when the queue of orders at the best bid or the best ask is depleted, or when an order arrives inside the spread. In the latent order book of [20 ] and in the continuum limits of [37 ] , the volume spreads over a continuum of prices, while the price remains the boundary between the bids and the asks, where volume is depleted. In our book the volume of each side also spreads over a continuum of distances, while no order reaches the mid price under Assumption 3.3(b), such that no volume is depleted at any price. As each desk keeps its side, the volume of each side moreover stays one, as normalized in § 2.1 . Each side is therefore a probability distribution over distances, and the mid price can move only through the shape of these two distributions near it. We now show how the two distributions evolve, how their evolution changes θ n \theta_n θ n , and how the motion of the mid price in turn reshapes the two distributions.
The mid price and the shape of the book. We first write the move of the mid price through the two distributions, such that its dependence on their shape becomes explicit. With the measure V n σ V_n^\sigma V n σ (§ 2.1 ), the volume of side σ \sigma σ between the distances a a a and b b b at step n n n is
V n σ ( ( a , b ] ) = ∑ i : σ i = σ m i ∑ j = 1 L 1 { x i , j ( n ) ∈ ( a , b ] } , 0 ≤ a < b . V_n^\sigma((a,b])=\sum_{i:\sigma_i=\sigma}m_i\sum_{j=1}^{L}\mathbf{1}\{x_{i,j}(n)\in(a,b]\},\qquad0\le a<b. V n σ (( a , b ]) = i : σ i = σ ∑ m i j = 1 ∑ L 1 { x i , j ( n ) ∈ ( a , b ]} , 0 ≤ a < b .
Then (1) and (2) read
Z n σ = ∫ 0 ∞ e − x / ℓ V n σ ( d x ) , θ n = − c ( Z n + − Z n − ) . Z_n^\sigma=\int_0^\infty e^{-x/\ell}V_n^\sigma(dx),\qquad\theta_n=-c\left(Z_n^+-Z_n^-\right). Z n σ = ∫ 0 ∞ e − x / ℓ V n σ ( d x ) , θ n = − c ( Z n + − Z n − ) .
such that Z n σ Z_n^\sigma Z n σ is the Laplace transform of the distribution of side σ \sigma σ at 1 / ℓ 1/\ell 1/ ℓ . One may picture the two sides as two fluids that press on the mid price from either side, each with the volume it holds near the mid price. The mid price then moves away from the side that presses harder. Let the asks and their targets follow, for instance, a Gamma distribution with shape 2 and scale 1.5, and let the bids and their targets be uniform on [ 1 , 11 ] [1,11] [ 1 , 11 ] . We then obtain
Z + = ( 1 + 1.5 ℓ ) − 2 , Z − = e − 1 / ℓ ℓ 10 ( 1 − e − 10 / ℓ ) , Z^+=\left(1+\frac{1.5}{\ell}\right)^{-2},\qquad Z^-=e^{-1/\ell}\frac{\ell}{10}\left(1-e^{-10/\ell}\right), Z + = ( 1 + ℓ 1.5 ) − 2 , Z − = e − 1/ ℓ 10 ℓ ( 1 − e − 10/ ℓ ) ,
such that Z + = 1 / 4 Z^+=1/4 Z + = 1/4 and Z − ≈ 0.077 Z^-\approx0.077 Z − ≈ 0.077 for ℓ = 1.5 \ell=1.5 ℓ = 1.5 . The asks thus press harder, and by (2) the mid price moves toward the bids with θ n ≈ − 0.17 c \theta_n\approx-0.17c θ n ≈ − 0.17 c .
The move of the mid price in turn reshapes both distributions. By (3) every ask moves away from the mid price by ∣ θ n ∣ \lvert\theta_n\rvert ∣ θ n ∣ , and every bid moves closer by ∣ θ n ∣ \lvert\theta_n\rvert ∣ θ n ∣ . As a shift of a whole side by Δ \Delta Δ multiplies its Laplace transform by e − Δ / ℓ e^{-\Delta/\ell} e − Δ/ ℓ , the shift lowers Z + Z^+ Z + and raises Z − Z^- Z − . Without the desks, the two sides would press equally after a total shift Δ ∗ \Delta^* Δ ∗ with
e − Δ ∗ / ℓ Z + = e Δ ∗ / ℓ Z − , Δ ∗ = ℓ 2 ln Z + Z − ≈ 0.88 , e^{-\Delta^*/\ell}Z^+=e^{\Delta^*/\ell}Z^-,\qquad\Delta^*=\frac{\ell}{2}\ln\frac{Z^+}{Z^-}\approx0.88, e − Δ ∗ / ℓ Z + = e Δ ∗ / ℓ Z − , Δ ∗ = 2 ℓ ln Z − Z + ≈ 0.88 ,
before any bid reaches the mid price. Yet each acting desk moves its orders back toward their targets, such that the asks again press harder. By the gradient rule (§ 2.4 ), the desks also move their orders with the trends P k P_k P k , which turn negative as the mid price falls. The asks then move closer and the bids away by Π ∣ P k ∣ \Pi\lvert P_k\rvert Π ∣ P k ∣ , and the mid price falls faster. The shape of the two distributions thus sets the move of the mid price, which reshapes the two distributions in turn. Throughout, θ n \theta_n θ n changes only because volume moves between distances, as the volume of each side stays one.
Flows in the protocol. In the protocol each step moves finitely many point masses, such that we can count the volume that moves between distances exactly. For a distance r ≥ 0 r\ge0 r ≥ 0 , let
Φ n σ ( r ) = ∑ i : σ i = σ m i ∑ j = 1 L ( 1 { x i , j ( n ) ≤ r < x i , j ( n + 1 ) } − 1 { x i , j ( n + 1 ) ≤ r < x i , j ( n ) } ) . \Phi_n^\sigma(r)=\sum_{i:\sigma_i=\sigma}m_i\sum_{j=1}^{L}\left(\mathbf{1}\{x_{i,j}(n)\le r<x_{i,j}(n+1)\}-\mathbf{1}\{x_{i,j}(n+1)\le r<x_{i,j}(n)\}\right). Φ n σ ( r ) = i : σ i = σ ∑ m i j = 1 ∑ L ( 1 { x i , j ( n ) ≤ r < x i , j ( n + 1 )} − 1 { x i , j ( n + 1 ) ≤ r < x i , j ( n )} ) .
denote the flow of side σ \sigma σ through r r r . Its first term is the volume of the orders that move across r r r away from the mid price in step n n n , while its second term is the volume of those that move across r r r toward it.
Lemma 3.12 (Flow identity of the protocol). For every side σ \sigma σ , every step n n n and all 0 ≤ a < b 0\le a<b 0 ≤ a < b ,
V n + 1 σ ( ( a , b ] ) = V n σ ( ( a , b ] ) + Φ n σ ( a ) − Φ n σ ( b ) . (13) V_{n+1}^\sigma((a,b])=V_n^\sigma((a,b])+\Phi_n^\sigma(a)-\Phi_n^\sigma(b). \tag*{(13)} V n + 1 σ (( a , b ]) = V n σ (( a , b ]) + Φ n σ ( a ) − Φ n σ ( b ) . (13)
Under Assumption 3.3 3.3 3.3 (b), moreover, Φ n σ ( 0 ) = 0 \Phi^{\sigma}_{n}(0)=0 Φ n σ ( 0 ) = 0 . Under the assumptions of Lemma 3.4 3.4 3.4 , Φ n σ ( r ) = 0 \Phi^{\sigma}_{n}(r)=0 Φ n σ ( r ) = 0 for n γ ≤ T n\gamma\le T nγ ≤ T and r > r T = max i , j x i , j i , j + R T r>r_T=\max_{i,j}x^{i,j}_{i,j}+R_T r > r T = max i , j x i , j i , j + R T .
The volume between a a a and b b b at the next step is thus its present volume, plus the flow into the interval across a a a , minus the flow out of it across b b b . In particular, no volume flows through the mid price or far from it.
Proof. For a single order with x = x i , j ( n ) x=x_{i,j}(n) x = x i , j ( n ) and x ′ = x i , j ( n + 1 ) x'=x_{i,j}(n+1) x ′ = x i , j ( n + 1 ) , we split the indicator of ( a , b ] (a,b] ( a , b ] at a a a and b b b ,
1 { x ′ ∈ ( a , b ] } − 1 { x ∈ ( a , b ] } = ( 1 { x ′ > a } − 1 { x > a } ) − ( 1 { x ′ > b } − 1 { x > b } ) , 1 { x ′ > r } − 1 { x > r } = 1 { x ≤ r < x ′ } − 1 { x ′ ≤ r < x } . \begin{aligned}
\mathbf{1}\{x'\in(a,b]\}-\mathbf{1}\{x\in(a,b]\}
&=\left(\mathbf{1}\{x'>a\}-\mathbf{1}\{x>a\}\right)-\left(\mathbf{1}\{x'>b\}-\mathbf{1}\{x>b\}\right),\\
\mathbf{1}\{x'>r\}-\mathbf{1}\{x>r\}
&=\mathbf{1}\{x\le r<x'\}-\mathbf{1}\{x'\le r<x\}.
\end{aligned} 1 { x ′ ∈ ( a , b ]} − 1 { x ∈ ( a , b ]} 1 { x ′ > r } − 1 { x > r } = ( 1 { x ′ > a } − 1 { x > a } ) − ( 1 { x ′ > b } − 1 { x > b } ) , = 1 { x ≤ r < x ′ } − 1 { x ′ ≤ r < x } .
Multiplying by m i m_i m i and summing over the orders of side σ \sigma σ , we obtain (13) . Under Assumption 3.3 3.3 3.3 (b) every order keeps x , x ′ > 0 x,x'>0 x , x ′ > 0 , such that both indicators in Φ n σ ( 0 ) \Phi^{\sigma}_{n}(0) Φ n σ ( 0 ) vanish. Lastly, from Lemma 3.4 3.4 3.4 we obtain x , x ′ ≤ r T x,x'\le r_T x , x ′ ≤ r T for n γ ≤ T n\gamma\le T nγ ≤ T . For r > r T r>r_T r > r T every indicator in Φ n σ ( r ) \Phi^{\sigma}_{n}(r) Φ n σ ( r ) therefore vanishes. □
Notice that the identity telescopes over adjacent intervals. For distances 0 ≤ r 0 < r 1 < ⋯ < r m 0\le r_0<r_1<\cdots<r_m 0 ≤ r 0 < r 1 < ⋯ < r m we obtain
∑ l = 1 m ( V n + 1 σ ( ( r l − 1 , r l ] ) − V n σ ( ( r l − 1 , r l ] ) ) = Φ n σ ( r 0 ) − Φ n σ ( r m ) . \sum_{l=1}^{m}\left(V^{\sigma}_{n+1}\left((r_{l-1},r_l]\right)-V^{\sigma}_{n}\left((r_{l-1},r_l]\right)\right)=\Phi^{\sigma}_{n}(r_0)-\Phi^{\sigma}_{n}(r_m). l = 1 ∑ m ( V n + 1 σ ( ( r l − 1 , r l ] ) − V n σ ( ( r l − 1 , r l ] ) ) = Φ n σ ( r 0 ) − Φ n σ ( r m ) .
such that volume leaving one interval through r l r_l r l enters the adjacent one. The identity (13) thus has the conservation form of finite-volume schemes [47 Section 4.1] , in which the quantity held in each interval changes only through the flows across its two endpoints. In deriving (13) we only relied on the observation that each step moves point masses. The identity therefore holds for any sampling, including the latent signals (Remark 3.11 ).
We now turn to the mid price. Since Z n σ Z^{\sigma}_{n} Z n σ weights every order by its distance from the mid price, we write this weight as an integral over the distances between the order and the mid price.
Proposition 3.13 (Motion of the mid price). Under Assumption 3.3 3.3 3.3 (b), for every step n n n ,
θ n + 1 − θ n = c ∫ 0 ∞ e − r / ℓ ℓ ( Φ n + ( r ) − Φ n − ( r ) ) d r . \theta_{n+1}-\theta_n=c\int_{0}^{\infty}\frac{e^{-r/\ell}}{\ell}\left(\Phi^{+}_{n}(r)-\Phi^{-}_{n}(r)\right)\,dr. θ n + 1 − θ n = c ∫ 0 ∞ ℓ e − r / ℓ ( Φ n + ( r ) − Φ n − ( r ) ) d r .
The move of the mid price thus responds only to flows, each weighted by e − r / ℓ e^{-r/\ell} e − r / ℓ according to its distance from the mid price. In the picture above, a side whose volume withdraws from the mid price presses less, such that the mid price moves toward it.
Proof. For x > 0 x>0 x > 0 we have
e − x / ℓ = ∫ 0 ∞ e − r / ℓ ℓ 1 { x ≤ r } d r . e^{-x/\ell}=\int_{0}^{\infty}\frac{e^{-r/\ell}}{\ell}\mathbf{1}\{x\le r\}\,dr. e − x / ℓ = ∫ 0 ∞ ℓ e − r / ℓ 1 { x ≤ r } d r .
As every order satisfies x > 0 x>0 x > 0 under Assumption 3.3 3.3 3.3 (b), we obtain
Z n σ = ∫ 0 ∞ e − r / ℓ ℓ V n σ ( ( 0 , r ] ) d r . Z^{\sigma}_{n}=\int_{0}^{\infty}\frac{e^{-r/\ell}}{\ell}V^{\sigma}_{n}((0,r])\,dr. Z n σ = ∫ 0 ∞ ℓ e − r / ℓ V n σ (( 0 , r ]) d r .
By Lemma 3.12 with Φ n σ ( 0 ) = 0 \Phi_n^\sigma(0)=0 Φ n σ ( 0 ) = 0 , the volume within r r r evolves as V n + 1 σ ( ( 0 , r ] ) = V n σ ( ( 0 , r ] ) − Φ n σ ( r ) V_{n+1}^\sigma((0,r]) = V_n^\sigma((0,r])-\Phi_n^\sigma(r) V n + 1 σ (( 0 , r ]) = V n σ (( 0 , r ]) − Φ n σ ( r ) . Hence
Z n + 1 σ − Z n σ = − ∫ 0 ∞ e − r / ℓ ℓ Φ n σ ( r ) d r . Z_{n+1}^\sigma-Z_n^\sigma=-\int_0^\infty\frac{e^{-r/\ell}}{\ell}\Phi_n^\sigma(r)\,dr. Z n + 1 σ − Z n σ = − ∫ 0 ∞ ℓ e − r / ℓ Φ n σ ( r ) d r .
and substituting into θ n + 1 − θ n = − c ( ( Z n + 1 + − Z n + ) − ( Z n + 1 − − Z n − ) ) \theta_{n+1}-\theta_n=-c((Z_{n+1}^+-Z_n^+)-(Z_{n+1}^--Z_n^-)) θ n + 1 − θ n = − c (( Z n + 1 + − Z n + ) − ( Z n + 1 − − Z n − )) proves the claim. □
Flows in the continuum. In the protocol each order enters or leaves an interval at most once per step. Lemma 3.12 thus counts every change of the volume of the interval. In the continuum, an order driven by Brownian noise instead flickers back and forth across each distance it crosses. By [53 Theorem 2.28] , it returns to that distance infinitely often near each crossing time, such that the crossings cannot be counted and the volume of an interval has no Itô differential. As for (1) in § 2.1 , we therefore count the volume beyond a distance with a smooth step of width ε \varepsilon ε rather than a sharp cutoff. An order within ε \varepsilon ε of the distance then counts only partly, and its share changes smoothly, such that Itô’s formula applies. To cover the SDE (Theorem 3.5 ) and its version with latent signals (Remark 3.11 ) at once, we let the distances move as
d x i , j = η i , j d t + d N i , j , d N i , j = ∑ l H ( i , j ) , l d W l . dx_{i,j}=\eta_{i,j}\,dt+dN_{i,j},\qquad dN_{i,j}=\sum_l H_{(i,j),l}\,dW_l. d x i , j = η i , j d t + d N i , j , d N i , j = l ∑ H ( i , j ) , l d W l .
where the W l W_l W l are finitely many independent Brownian motions and the coefficients η i , j \eta_{i,j} η i , j and H ( i , j ) , l H_{(i,j),l} H ( i , j ) , l are bounded and adapted. The noise of two orders then has the covariance rate Q ( i , j ) , ( i ′ , j ′ ) = ∑ l H ( i , j ) , l H ( i ′ , j ′ ) , l Q_{(i,j),(i',j')}=\sum_l H_{(i,j),l}H_{(i',j'),l} Q ( i , j ) , ( i ′ , j ′ ) = ∑ l H ( i , j ) , l H ( i ′ , j ′ ) , l , and we write A i , j = Q ( i , j ) , ( i , j ) A_{i,j}=Q_{(i,j),(i,j)} A i , j = Q ( i , j ) , ( i , j ) . For the SDE (Theorem 3.5 ), since x i , j = σ i y i , j x_{i,j}=\sigma_i y_{i,j} x i , j = σ i y i , j , we have η i , j = σ i V i , j \eta_{i,j}=\sigma_i V_{i,j} η i , j = σ i V i , j with V i , j V_{i,j} V i , j (Proposition 3.10 ), and the Brownian motions are the W i W_i W i of the desks with H ( i , j ) , l = σ i γ G i , j 1 { l = i } H_{(i,j),l}=\sigma_i\sqrt{\gamma}\,G_{i,j}\mathbf{1}\{l=i\} H ( i , j ) , l = σ i γ G i , j 1 { l = i } .
Let ρ \rho ρ be a smooth even function with support in [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] , with ρ ≥ 0 \rho\geq0 ρ ≥ 0 and ∫ ρ = 1 \int\rho=1 ∫ ρ = 1 , and let
ρ ε ( z ) = 1 ε ρ ( z ε ) , ψ ε ( z ) = ∫ − ∞ z ρ ε ( u ) d u . \rho_\varepsilon(z)=\frac{1}{\varepsilon}\rho\left(\frac{z}{\varepsilon}\right),\qquad\psi_\varepsilon(z)=\int_{-\infty}^{z}\rho_\varepsilon(u)\,du. ρ ε ( z ) = ε 1 ρ ( ε z ) , ψ ε ( z ) = ∫ − ∞ z ρ ε ( u ) d u .
The function ψ ε \psi_\varepsilon ψ ε rises smoothly from zero to one on [ − ε , ε ] [-\varepsilon,\varepsilon] [ − ε , ε ] . The sum ∑ i : σ i = σ m i ∑ j = 1 L ψ ε ( x i , j − r ) \sum_{i:\sigma_i=\sigma}m_i\sum_{j=1}^L\psi_\varepsilon(x_{i,j}-r) ∑ i : σ i = σ m i ∑ j = 1 L ψ ε ( x i , j − r ) is therefore the volume of side σ \sigma σ beyond r r r , smoothed over the width ε \varepsilon ε . We then define the smoothed flow of side σ \sigma σ through r r r by
d Φ t ε , σ ( r ) = ∑ i : σ i = σ m i ∑ j = 1 L ( ρ ε ( x i , j − r ) ( η i , j d t + d N i , j ) + 1 2 ρ ε ′ ( x i , j − r ) A i , j d t ) . (14) d\Phi_t^{\varepsilon,\sigma}(r)=\sum_{i:\sigma_i=\sigma}m_i\sum_{j=1}^{L}\left(\rho_\varepsilon(x_{i,j}-r)(\eta_{i,j}\,dt+dN_{i,j})+\frac{1}{2}\rho_\varepsilon'(x_{i,j}-r)A_{i,j}\,dt\right). \tag*{(14)} d Φ t ε , σ ( r ) = i : σ i = σ ∑ m i j = 1 ∑ L ( ρ ε ( x i , j − r ) ( η i , j d t + d N i , j ) + 2 1 ρ ε ′ ( x i , j − r ) A i , j d t ) . (14)
Proposition 3.14 (Flow identity in the continuum). For every side σ \sigma σ , every ε > 0 \varepsilon>0 ε > 0 and all 0 ≤ a < b 0\leq a<b 0 ≤ a < b , the smoothed indicator φ a , b ε = ψ ε ( ⋅ − a ) − ψ ε ( ⋅ − b ) \varphi_{a,b}^\varepsilon=\psi_\varepsilon(\cdot-a)-\psi_\varepsilon(\cdot-b) φ a , b ε = ψ ε ( ⋅ − a ) − ψ ε ( ⋅ − b ) of ( a , b ] (a,b] ( a , b ] satisfies
d ∑ i : σ i = σ m i ∑ j = 1 L φ a , b ε ( x i , j ) = d Φ t ε , σ ( a ) − d Φ t ε , σ ( b ) , d\sum_{i:\sigma_i=\sigma}m_i\sum_{j=1}^{L}\varphi_{a,b}^\varepsilon(x_{i,j})=d\Phi_t^{\varepsilon,\sigma}(a)-d\Phi_t^{\varepsilon,\sigma}(b), d i : σ i = σ ∑ m i j = 1 ∑ L φ a , b ε ( x i , j ) = d Φ t ε , σ ( a ) − d Φ t ε , σ ( b ) , and the smoothed flow (14) reads
d Φ t ε , σ ( r ) = ( ∑ i : σ i = σ m i ∑ j = 1 L ρ ε ( x i , j − r ) η i , j − 1 2 ∂ r ∑ i : σ i = σ m i ∑ j = 1 L ρ ε ( x i , j − r ) A i , j ) d t + ∑ i : σ i = σ m i ∑ j = 1 L ρ ε ( x i , j − r ) d N i , j . d\Phi_t^{\varepsilon,\sigma}(r)=\left(\sum_{i:\sigma_i=\sigma}m_i\sum_{j=1}^{L}\rho_\varepsilon(x_{i,j}-r)\eta_{i,j}-\frac{1}{2}\partial_r\sum_{i:\sigma_i=\sigma}m_i\sum_{j=1}^{L}\rho_\varepsilon(x_{i,j}-r)A_{i,j}\right)dt
+\sum_{i:\sigma_i=\sigma}m_i\sum_{j=1}^{L}\rho_\varepsilon(x_{i,j}-r)\,dN_{i,j}. d Φ t ε , σ ( r ) = ( i : σ i = σ ∑ m i j = 1 ∑ L ρ ε ( x i , j − r ) η i , j − 2 1 ∂ r i : σ i = σ ∑ m i j = 1 ∑ L ρ ε ( x i , j − r ) A i , j ) d t + i : σ i = σ ∑ m i j = 1 ∑ L ρ ε ( x i , j − r ) d N i , j .
The smoothed flow through r r r thus depends only on the orders within ε \varepsilon ε of r r r .
Proof. Since ψ ε ′ = ρ ε \psi_\varepsilon'=\rho_\varepsilon ψ ε ′ = ρ ε and ψ ε ′ ′ = ρ ε ′ \psi_\varepsilon''=\rho_\varepsilon' ψ ε ′′ = ρ ε ′ , we obtain from Itô’s formula [RY99, Theorem IV.3.3] for each order
d ψ ε ( x i , j − r ) = ρ ε ( x i , j − r ) ( η i , j d t + d N i , j ) + 1 2 ρ ε ′ ( x i , j − r ) A i , j d t . d\psi_\varepsilon(x_{i,j}-r)=\rho_\varepsilon(x_{i,j}-r)(\eta_{i,j}\,dt+dN_{i,j})+\frac{1}{2}\rho_\varepsilon'(x_{i,j}-r)A_{i,j}\,dt. d ψ ε ( x i , j − r ) = ρ ε ( x i , j − r ) ( η i , j d t + d N i , j ) + 2 1 ρ ε ′ ( x i , j − r ) A i , j d t .
Multiplying by m i m_i m i and summing over the orders of side σ \sigma σ , we obtain d Φ t ε , σ ( r ) d\Phi_t^{\varepsilon,\sigma}(r) d Φ t ε , σ ( r ) . As Φ t ε \Phi_{t}^{\varepsilon} Φ t ε is the difference of two such steps, the identity for ( a , b ] (a,b] ( a , b ] follows. Lastly, since ∂ r ρ ε ( x − r ) = − ρ ε ′ ( x − r ) \partial_r\rho_\varepsilon(x-r)=-\rho_\varepsilon'(x-r) ∂ r ρ ε ( x − r ) = − ρ ε ′ ( x − r ) , we obtain
1 2 ∑ i : σ i = σ m i ∑ j = 1 L ρ ε ′ ( x i , j − r ) A i , j = − 1 2 ∂ r ∑ i : σ i = σ m i ∑ j = 1 L ρ ε ( x i , j − r ) A i , j , \frac{1}{2}\sum_{i:\sigma_i=\sigma}m_i\sum_{j=1}^{L}\rho_\varepsilon'(x_{i,j}-r)A_{i,j}=-\frac{1}{2}\partial_r\sum_{i:\sigma_i=\sigma}m_i\sum_{j=1}^{L}\rho_\varepsilon(x_{i,j}-r)A_{i,j}, 2 1 i : σ i = σ ∑ m i j = 1 ∑ L ρ ε ′ ( x i , j − r ) A i , j = − 2 1 ∂ r i : σ i = σ ∑ m i j = 1 ∑ L ρ ε ( x i , j − r ) A i , j ,
which is the second form of the smoothed flow.
As in the protocol, the identity telescopes over adjacent intervals, such that volume leaving one interval enters the adjacent one for every width ε \varepsilon ε .
No volume crosses the mid price. In the protocol no volume flows through the mid price or far from it, while the noise of the SDE can carry an order across the mid price. By Theorem 3.5 , however, the SDE stays within O ( γ ) O(\sqrt{\gamma}) O ( γ ) of the solution X ‾ \overline{X} X of the ODE d X ‾ = F ~ ( X ‾ ) d t d\overline{X}=\widetilde{F}(\overline{X})\,dt d X = F ( X ) d t . We show that a crossing therefore occurs only with probability O ( γ ) O(\gamma) O ( γ ) whenever X ‾ \overline{X} X keeps its orders away from the mid price. Let x ‾ i , j = σ i y ‾ i , j \overline{x}_{i,j}=\sigma_i\overline{y}_{i,j} x i , j = σ i y i , j denote the distances under X ‾ \overline{X} X , and let r ‾ T = sup t ≤ T max i , j x ‾ i , j ( t ) \overline{r}_T=\sup_{t\leq T}\max_{i,j}\overline{x}_{i,j}(t) r T = sup t ≤ T max i , j x i , j ( t ) .
Corollary 3.15 (No volume crosses the mid price). Let the assumptions of Theorem 3.5 hold, and let δ > 0 \delta>0 δ > 0 satisfy x ‾ i , j ( t ) ≥ 2 δ \overline{x}_{i,j}(t)\geq2\delta x i , j ( t ) ≥ 2 δ for every order and t ≤ T t\leq T t ≤ T . Then the event E δ = { sup t ≤ T ∣ X ( t ) − X ‾ ( t ) ∣ < δ } E_\delta=\{\sup_{t\leq T}|X(t)-\overline{X}(t)|<\delta\} E δ = { sup t ≤ T ∣ X ( t ) − X ( t ) ∣ < δ } satisfies P ( E δ c ) ≤ C γ / δ 2 \mathbb{P}(E_\delta^{c})\leq C\gamma/\delta^2 P ( E δ c ) ≤ C γ / δ 2 . On E δ E_\delta E δ , for every side σ \sigma σ , every ε < δ \varepsilon<\delta ε < δ and t ≤ T t\leq T t ≤ T , (i) every order satisfies x i , j > δ x_{i,j}>\delta x i , j > δ , and the volume of each side stays one; (ii) d Φ t ε , σ ( r ) = 0 d\Phi_t^{\varepsilon,\sigma}(r)=0 d Φ t ε , σ ( r ) = 0 for r ≥ r ‾ T + δ + ε r\geq\overline{r}_T+\delta+\varepsilon r ≥ r T + δ + ε .
Up to an event of probability O ( γ ) O(\gamma) O ( γ ) , the SDE thus keeps every order away from the mid price and lets no volume flow far from it.
Proof. By Chebyshev’s inequality and Theorem 3.5 (ii), we obtain
P ( E δ c ) ≤ 1 δ 2 E sup t ≤ T ∣ X ( t ) − X ‾ ( t ) ∣ 2 ≤ C γ δ 2 . \mathbb{P}(E_\delta^{c})\leq\frac{1}{\delta^2}\mathbb{E}\sup_{t\leq T}|X(t)-\overline{X}(t)|^2\leq\frac{C\gamma}{\delta^2}. P ( E δ c ) ≤ δ 2 1 E t ≤ T sup ∣ X ( t ) − X ( t ) ∣ 2 ≤ δ 2 C γ .
On E δ E_\delta E δ every order satisfies ∣ x i , j − x ‾ i , j ∣ < δ |x_{i,j}-\overline{x}_{i,j}|<\delta ∣ x i , j − x i , j ∣ < δ for t ≤ T t\leq T t ≤ T , such that
x i , j > x ‾ i , j − δ ≥ δ , x i , j < r ‾ T + δ . x_{i,j}>\overline{x}_{i,j}-\delta\geq\delta,\qquad x_{i,j}<\overline{r}_T+\delta. x i , j > x i , j − δ ≥ δ , x i , j < r T + δ .
Every order therefore keeps its side, such that the volume of each side stays ∑ i : σ i = σ L m i = 1 \sum_{i:\sigma_i=\sigma}Lm_i=1 ∑ i : σ i = σ L m i = 1 . Further, every order stays farther than ε \varepsilon ε from each r ≥ r ‾ T + δ + ε r\geq\overline{r}_T+\delta+\varepsilon r ≥ r T + δ + ε . At these distances every term of (14) carries the factor ρ ε ( x i , j − r ) = 0 \rho_\varepsilon(x_{i,j}-r)=0 ρ ε ( x i , j − r ) = 0 , such that the drift of the smoothed flow vanishes on E δ E_\delta E δ . Its noise is a continuous martingale with quadratic variation
∫ 0 t ∑ ( i , j ) , ( i ′ , j ′ ) m i m i ′ ρ ε ( x i , j − r ) ρ ε ( x i ′ , j ′ − r ) Q ( i , j ) , ( i ′ , j ′ ) d s , \int_0^t\sum_{(i,j),(i',j')}m_im_{i'}\rho_\varepsilon(x_{i,j}-r)\rho_\varepsilon(x_{i',j'}-r)Q_{(i,j),(i',j')}\,ds, ∫ 0 t ( i , j ) , ( i ′ , j ′ ) ∑ m i m i ′ ρ ε ( x i , j − r ) ρ ε ( x i ′ , j ′ − r ) Q ( i , j ) , ( i ′ , j ′ ) d s ,
which vanishes on E δ E_\delta E δ . By [RY99, Proposition IV.1.13], a continuous martingale is constant wherever its quadratic variation is constant, hence d Φ t ε , σ ( r ) = 0 d\Phi_t^{\varepsilon,\sigma}(r)=0 d Φ t ε , σ ( r ) = 0 at these distances. □
The mid price in the continuum. We now show that Proposition 3.13 holds in the continuum as well. On the event E δ E_\delta E δ every order stays farther than ε \varepsilon ε from the mid price, such that smoothing the distances over the width ε \varepsilon ε rescales every weight e − x / ℓ e^{-x/\ell} e − x / ℓ by the same constant
κ ε = ∫ ρ ε ( z ) e − z / ℓ d z = 1 + O ( ε 2 ℓ 2 ) . \kappa_\varepsilon=\int\rho_\varepsilon(z)e^{-z/\ell}\,dz=1+O\left(\frac{\varepsilon^2}{\ell^2}\right). κ ε = ∫ ρ ε ( z ) e − z / ℓ d z = 1 + O ( ℓ 2 ε 2 ) .
Proposition 3.16 (Motion of the mid price in the continuum). Under the assumptions of Corollary 3.15 , on the event E δ E_\delta E δ , for every ε < δ \varepsilon<\delta ε < δ and t ≤ T t\leq T t ≤ T ,
d v t = c ˉ κ ε ∫ 0 ∞ e − r / ℓ ℓ ( d Φ t ε , + ( r ) − d Φ t ε , − ( r ) ) d r . dv_t=\frac{\bar{c}}{\kappa_\varepsilon}\int_0^\infty\frac{e^{-r/\ell}}{\ell}\left(d\Phi_t^{\varepsilon,+}(r)-d\Phi_t^{\varepsilon,-}(r)\right)\,dr. d v t = κ ε c ˉ ∫ 0 ∞ ℓ e − r / ℓ ( d Φ t ε , + ( r ) − d Φ t ε , − ( r ) ) d r .
The velocity of the mid price thus responds only to the smoothed flows, with the weights (Proposition 3.13 ) up to the factor κ ε \kappa_\varepsilon κ ε , which tends to one as ε → 0 \varepsilon\to0 ε → 0 .
Proof. For x ≥ ε x\geq\varepsilon x ≥ ε , every z z z in the support of ρ ε \rho_\varepsilon ρ ε satisfies x + z ≥ 0 x+z\geq0 x + z ≥ 0 . Writing 1 − ψ ε ( x − r ) = ψ ε ( r − x ) 1-\psi_\varepsilon(x-r)=\psi_\varepsilon(r-x) 1 − ψ ε ( x − r ) = ψ ε ( r − x ) , which holds as ρ \rho ρ is even, we obtain
∫ 0 ∞ e − r / ℓ ℓ ψ ε ( r − x ) d r = ∫ ρ ε ( z ) ∫ x + z ∞ e − r / ℓ ℓ d r d z = ∫ ρ ε ( z ) e − ( x + z ) / ℓ d z = κ ε e − x / ℓ . \int_0^\infty\frac{e^{-r/\ell}}{\ell}\psi_\varepsilon(r-x)\,dr=\int\rho_\varepsilon(z)\int_{x+z}^\infty\frac{e^{-r/\ell}}{\ell}\,dr\,dz=\int\rho_\varepsilon(z)e^{-(x+z)/\ell}\,dz=\kappa_\varepsilon e^{-x/\ell}. ∫ 0 ∞ ℓ e − r / ℓ ψ ε ( r − x ) d r = ∫ ρ ε ( z ) ∫ x + z ∞ ℓ e − r / ℓ d r d z = ∫ ρ ε ( z ) e − ( x + z ) / ℓ d z = κ ε e − x / ℓ .
On E δ E_\delta E δ every order satisfies x i , j > δ > ε x_{i,j}>\delta>\varepsilon x i , j > δ > ε . Summing over the orders of side σ \sigma σ with ∑ i : σ i = σ L m i = 1 \sum_{i:\sigma_i=\sigma}Lm_i=1 ∑ i : σ i = σ L m i = 1 , we therefore obtain
κ ε Z t σ = ∫ 0 ∞ e − r / ℓ ℓ ( 1 − ∑ i : σ i = σ m i ∑ j = 1 L ψ ε ( x i , j − r ) ) d r . \kappa_\varepsilon Z_t^\sigma=\int_0^\infty\frac{e^{-r/\ell}}{\ell}\left(1-\sum_{i:\sigma_i=\sigma}m_i\sum_{j=1}^{L}\psi_\varepsilon(x_{i,j}-r)\right)\,dr. κ ε Z t σ = ∫ 0 ∞ ℓ e − r / ℓ ( 1 − i : σ i = σ ∑ m i j = 1 ∑ L ψ ε ( x i , j − r ) ) d r .
The sum in the integrand is the smoothed volume of side σ \sigma σ beyond r r r , whose differential is d Φ t ε , σ ( r ) d\Phi_t^{\varepsilon,\sigma}(r) d Φ t ε , σ ( r ) by the proof of Proposition 3.14 . Since the integrand is bounded and continuous in r r r , and ℓ − 1 e − r / ℓ d r \ell^{-1}e^{-r/\ell}\,dr ℓ − 1 e − r / ℓ d r is a probability measure, the stochastic Fubini theorem [Ver12, Theorem 2.2] allows us to interchange the integral over r r r with the stochastic integrals, and we obtain
κ ε d Z t σ = − ∫ 0 ∞ e − r / ℓ ℓ d Φ t ε , σ ( r ) d r . \kappa_\varepsilon dZ_t^\sigma=-\int_0^\infty\frac{e^{-r/\ell}}{\ell}\,d\Phi_t^{\varepsilon,\sigma}(r)\,dr. κ ε d Z t σ = − ∫ 0 ∞ ℓ e − r / ℓ d Φ t ε , σ ( r ) d r .
Substituting into v t = − c ˉ ( Z t + − Z t − ) v_t=-\bar{c}(Z_t^+-Z_t^-) v t = − c ˉ ( Z t + − Z t − ) proves the claim. □
Feedback of the mid price. The shape of the two distributions sets the velocity v v v of the mid price, while the motion of the mid price moves every order through the shift and through the trends π k \pi_k π k . We now write this loop as one equation for v v v . To this end we apply Itô’s formula to v = − c ˉ ∑ i σ i m i ∑ j e − x i , j / ℓ v=-\bar{c}\sum_i\sigma_i m_i\sum_j e^{-x_{i,j}/\ell} v = − c ˉ ∑ i σ i m i ∑ j e − x i , j / ℓ directly, and we then split the drift of the distances into its causes.
Proposition 3.17 (Feedback of the mid price). For distances that move as in Proposition 3.14 ,
d v t = − c ˉ ℓ ∑ i σ i m i ∑ j = 1 L e − x i , j / ℓ ( η i , j d t + d N i , j − 1 2 ℓ A i , j d t ) . dv_t=-\frac{\bar{c}}{\ell}\sum_i\sigma_i m_i\sum_{j=1}^{L}e^{-x_{i,j}/\ell}\left(\eta_{i,j}\,dt+dN_{i,j}-\frac{1}{2\ell}A_{i,j}\,dt\right). d v t = − ℓ c ˉ i ∑ σ i m i j = 1 ∑ L e − x i , j / ℓ ( η i , j d t + d N i , j − 2 ℓ 1 A i , j d t ) . :::`
For the SDE (Theorem 3.5 ) with the state in K T K_T K T , the drift of v v v splits as
c ˉ ℓ ( − ( Z t + + Z t − ) v t + Π ∑ k = 1 K p k π k Z k , t ) + ∑ i σ i p k ( i ) m i ∑ j = 1 L e − x i , j / ℓ ( x i , j ∗ − x i , j ) − 1 2 ℓ ∑ i σ i m i ∑ j = 1 L e − x i , j / ℓ A i , j , \frac{\bar{c}}{\ell}\left(-(Z_t^{+}+Z_t^{-})v_t+\Pi\sum_{k=1}^{K}p_k\pi_kZ_{k,t}\right)
+\sum_i\sigma_ip_{k(i)}m_i\sum_{j=1}^{L}e^{-x_{i,j}/\ell}\left(x_{i,j}^{*}-x_{i,j}\right)
-\frac{1}{2\ell}\sum_i\sigma_im_i\sum_{j=1}^{L}e^{-x_{i,j}/\ell}A_{i,j}, ℓ c ˉ ( − ( Z t + + Z t − ) v t + Π k = 1 ∑ K p k π k Z k , t ) + i ∑ σ i p k ( i ) m i j = 1 ∑ L e − x i , j / ℓ ( x i , j ∗ − x i , j ) − 2 ℓ 1 i ∑ σ i m i j = 1 ∑ L e − x i , j / ℓ A i , j ,
where Z k , t = ∑ i : k ( i ) = k m i ∑ j e − x i , j / ℓ Z_{k,t}=\sum_{i:k(i)=k}m_i\sum_j e^{-x_{i,j}/\ell} Z k , t = ∑ i : k ( i ) = k m i ∑ j e − x i , j / ℓ is the volume of cluster k k k near the mid price.
The first term restores the balance of the two sides. A rising mid price shifts the asks toward it and the bids away from it, such that the asks press harder and v v v relaxes at the rate c ˉ ( Z + + Z − ) / ℓ \bar{c}(Z^{+}+Z^{-})/\ell c ˉ ( Z + + Z − ) / ℓ . The second term reinforces the motion through the memory of the desks. A positive trend π k \pi_k π k makes the desks of cluster k k k move their asks away and their bids closer, which raises v v v . The trend in turn follows v v v through d π k = ( v − π k ) d t / τ k d\pi_k=(v-\pi_k)\,dt/\tau_k d π k = ( v − π k ) d t / τ k . The third term pulls each side toward the shape of its targets, with a weight set by the probability of acting p k p_k p k . The fourth term arises from the fluctuations of the orders, as the convexity of e − x / ℓ e^{-x/\ell} e − x / ℓ makes fluctuating orders press harder on average. Lastly, the noise term carries the random timing of the desks near the mid price into v v v . The shape of the two distributions thus sets v v v . In turn, v v v reshapes the two distributions through the shift and the trends, and the reshaped distributions set the next velocity.
Proof. Itô’s formula for each distance gives
d e − x i , j / ℓ = − 1 ℓ e − x i , j / ℓ ( η i , j d t + d N i , j ) + 1 2 ℓ 2 e − x i , j / ℓ A i , j d t . de^{-x_{i,j}/\ell}=-\frac{1}{\ell}e^{-x_{i,j}/\ell}\left(\eta_{i,j}\,dt+dN_{i,j}\right)+\frac{1}{2\ell^2}e^{-x_{i,j}/\ell}A_{i,j}\,dt. d e − x i , j / ℓ = − ℓ 1 e − x i , j / ℓ ( η i , j d t + d N i , j ) + 2 ℓ 2 1 e − x i , j / ℓ A i , j d t .
Multiplying by − c ˉ σ i m i -\bar{c}\sigma_im_i − c ˉ σ i m i and summing over all orders, we obtain the first display. For the SDE (Theorem 3.5 ) on K T K_T K T , the drift of each distance reads
η i , j = σ i ( p k ( i ) b i , j − v ) = p k ( i ) ( x i , j ∗ − x i , j ) + p k ( i ) σ i Π π k ( i ) − σ i v . \eta_{i,j}=\sigma_i\left(p_{k(i)}b_{i,j}-v\right)=p_{k(i)}\left(x_{i,j}^{*}-x_{i,j}\right)+p_{k(i)}\sigma_i\Pi\pi_{k(i)}-\sigma_iv. η i , j = σ i ( p k ( i ) b i , j − v ) = p k ( i ) ( x i , j ∗ − x i , j ) + p k ( i ) σ i Π π k ( i ) − σ i v .
since σ i b i , j = − ( x i , j − x i , j ∗ ) + σ i Π π k ( i ) \sigma_ib_{i,j}=-(x_{i,j}-x_{i,j}^{*})+\sigma_i\Pi\pi_{k(i)} σ i b i , j = − ( x i , j − x i , j ∗ ) + σ i Π π k ( i ) . Substituting and using σ i 2 = 1 \sigma_i^2=1 σ i 2 = 1 , the term − σ i v -\sigma_iv − σ i v contributes − ( Z + + Z − ) v -(Z^{+}+Z^{-})v − ( Z + + Z − ) v and the term p k ( i ) σ i Π π k ( i ) p_{k(i)}\sigma_i\Pi\pi_{k(i)} p k ( i ) σ i Π π k ( i ) contributes Π ∑ k p k π k Z k \Pi\sum_k p_k\pi_kZ_k Π ∑ k p k π k Z k , which proves the split. □
The two distributions thus move the mid price only through flows of volume, and Proposition 3.17 closes the loop between their shape and the velocity v v v . Its noise term comes from the random timing of the desks, whose probabilities of acting p k p_k p k grow with the estimates U k U_k U k of market activity by (8) . Every action enters I n I_n I n in (7) and through (5) the estimates U k U_k U k , such that each action can raise the probability that other desks act in later steps. We study this self-excitation of the order flow next.
From the order book to self-exciting order flow
Linear response
In Section 3.3 we found that the noise of the mid price comes from the random timing of the desks. Yet this timing depends on the past. By (8) each desk acts with a probability that grows with its estimate U k U_k U k of market activity, while by (5) this estimate is a moving average of the size of the updates I n I_n I n in (7) . The order flow thus excites itself, as measured in many order books [25 , 35 ] . Roughness requires that the order flow excites itself strongly, with a memory that fades as a power law [22 , 41 ] . Models of order flow usually choose this memory to fit the data [9 ] . In our market, (5) and (8) instead fix the horizons over which the desks average the size of the updates, while the book fixes how large each update is. We can therefore compute the memory of the order flow from the desks. We first state the structure that this memory takes.
Self-exciting sequences. To make this memory precise, we use the structure of a Hawkes process [34 ] in discrete time [45 ] . Unlike a Poisson process, whose events occur at a constant rate, a Hawkes process lets each event raise the rate of later events. As the size of the updates takes real values rather than counting events, we state this structure for real values.
Definition 4.1 (Self-exciting sequence). Let ( Y n ) n ≥ 0 (Y_n)_{n\geq0} ( Y n ) n ≥ 0 be a stochastic process with values in [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) , such that Y n Y_n Y n is F n + 1 \mathcal{F}_{n+1} F n + 1 -measurable for the σ \sigma σ -algebra F n \mathcal{F}_n F n generated by ξ ( 0 ) , … , ξ ( n − 1 ) \xi(0),\ldots,\xi(n-1) ξ ( 0 ) , … , ξ ( n − 1 ) (§ 3.2 ). We call ( Y n ) (Y_n) ( Y n ) self-exciting with base rate μ ≥ 0 \mu\geq0 μ ≥ 0 and kernel ϕ : N ∖ { 0 } → [ 0 , ∞ ) \phi:\mathbb{N}\setminus\{0\}\to[0,\infty) ϕ : N ∖ { 0 } → [ 0 , ∞ ) with ∥ ϕ ∥ = ∑ s ∈ N ∖ { 0 } ϕ ( s ) < ∞ \lVert\phi\rVert=\sum_{s\in\mathbb{N}\setminus\{0\}}\phi(s)<\infty ∥ ϕ ∥ = ∑ s ∈ N ∖ { 0 } ϕ ( s ) < ∞ if, for every step n n n ,
E [ Y n ∣ F n ] = μ + ∑ m = 0 n − 1 ϕ ( n − m ) Y m . \mathbb{E}[Y_n\mid\mathcal{F}_n]=\mu+\sum_{m=0}^{n-1}\phi(n-m)Y_m. E [ Y n ∣ F n ] = μ + m = 0 ∑ n − 1 ϕ ( n − m ) Y m .
With ϕ = 0 \phi=0 ϕ = 0 the expected value stays at μ \mu μ , as for a Poisson process. With ϕ > 0 \phi>0 ϕ > 0 , each value Y m Y_m Y m instead raises the expected value of Y m + s Y_{m+s} Y m + s by ϕ ( s ) Y m \phi(s)Y_m ϕ ( s ) Y m , and of all later values together by ∥ ϕ ∥ Y m \lVert\phi\rVert Y_m ∥ ϕ ∥ Y m . The values raised in this way raise later values in turn. Repeating this argument, Y m > 0 Y_m>0 Y m > 0 raises all later values together in expectation by
Y m ( ∥ ϕ ∥ + ∥ ϕ ∥ 2 + ∥ ϕ ∥ 3 + ⋯ ) = { Y m ∥ ϕ ∥ 1 − ∥ ϕ ∥ , ∥ ϕ ∥ < 1 , ∞ , ∥ ϕ ∥ ≥ 1. Y_m\left(\lVert\phi\rVert+\lVert\phi\rVert^2+\lVert\phi\rVert^3+\cdots\right)=
\begin{cases}
Y_m\frac{\lVert\phi\rVert}{1-\lVert\phi\rVert}, & \lVert\phi\rVert<1,\\
\infty, & \lVert\phi\rVert\geq1.
\end{cases} Y m ( ∥ ϕ ∥ + ∥ ϕ ∥ 2 + ∥ ϕ ∥ 3 + ⋯ ) = { Y m 1 − ∥ ϕ ∥ ∥ ϕ ∥ , ∞ , ∥ ϕ ∥ < 1 , ∥ ϕ ∥ ≥ 1.
Each value thus raises later values by a finite amount on average if ∥ ϕ ∥ < 1 \lVert\phi\rVert<1 ∥ ϕ ∥ < 1 , while otherwise the values grow without bound.
Activity. The size of the updates grows with the number of desks that act in a step. A desk with more volume moves more of the book. We therefore count each desk by its volume. Let
A n = ∑ k = 1 K w k p k ( n ) \mathcal{A}_n=\sum_{k=1}^{K}w_k p_k(n) A n = k = 1 ∑ K w k p k ( n )
denote the activity in step n n n , where p k ( n ) p_k(n) p k ( n ) is the probability of acting (8) of cluster k k k . As the desks of cluster k k k on each side together hold the volume w k w_k w k by § 2.3 , the activity is the expected volume of the desks that act in step n n n on each side.
Size of the updates per unit activity. The size of the updates of an acting desk depends on where its orders sit, as a step deep in the book changes Z ∘ Z^\circ Z ∘ little. Let Δ Z i ( n ) \Delta Z_i(n) Δ Z i ( n ) denote the change Δ Z i \Delta Z_i Δ Z i (§ 2.4 ) that desk d i d_i d i causes if it acts in step n n n , and let
κ l ( n ) = c w l ∑ i : k ( i ) = l m i ∣ Δ Z i ( n ) ∣ \kappa_l(n) = \frac{c}{w_l} \sum_{i:k(i)=l} m_i \lvert\Delta Z_i(n) \rvert κ l ( n ) = w l c i : k ( i ) = l ∑ m i ∣ Δ Z i ( n )∣
denote the size of the updates per unit activity of cluster l l l .
Lemma 4.2 (Size of the updates per unit activity). For every step n n n ,
E [ I n ∣ F n ] = ∑ l = 1 K κ l ( n ) w l p l ( n ) . \mathbb{E}[I_n \mid\mathcal{F}_n] = \sum_{l=1}^{K} \kappa_l(n) w_l p_l(n). E [ I n ∣ F n ] = l = 1 ∑ K κ l ( n ) w l p l ( n ) .
The expected size of the updates thus grows with the activity of every cluster, at a rate κ l \kappa_l κ l that is large when the orders of cluster l l l sit near the mid price.
Proof. By § 2.4 , Δ Z i ( n ) \Delta Z_i(n) Δ Z i ( n ) depends only on the distances of the orders of d i d_i d i before its step and on the trend P k ( i ) P_{k(i)} P k ( i ) . Both are part of the state X n X_n X n , such that Δ Z i ( n ) \Delta Z_i(n) Δ Z i ( n ) is F n \mathcal{F}_n F n -measurable. Starting from (7) and using E [ ξ i ( n ) ∣ F n ] = p k ( i ) ( n ) \mathbb{E}[\xi_i(n) \mid\mathcal{F}_n] = p_{k(i)}(n) E [ ξ i ( n ) ∣ F n ] = p k ( i ) ( n ) , we obtain
E [ I n ∣ F n ] = ∑ i E [ ξ i ( n ) ∣ F n ] c m i ∣ Δ Z i ( n ) ∣ = ∑ l = 1 K p l ( n ) ∑ i : k ( i ) = l c m i ∣ Δ Z i ( n ) ∣ = ∑ l = 1 K κ l ( n ) w l p l ( n ) , \begin{aligned}
\mathbb{E}[I_n \mid\mathcal{F}_n] = \sum_i \mathbb{E}[\xi_i(n) \mid\mathcal{F}_n] c m_i \lvert\Delta Z_i(n) \rvert\\
&= \sum_{l=1}^{K} p_l(n) \sum_{i:k(i)=l} c m_i \lvert\Delta Z_i(n) \rvert= \sum_{l=1}^{K} \kappa_l(n) w_l p_l(n),
\end{aligned} E [ I n ∣ F n ] = i ∑ E [ ξ i ( n ) ∣ F n ] c m i ∣ Δ Z i ( n )∣ = l = 1 ∑ K p l ( n ) i : k ( i ) = l ∑ c m i ∣ Δ Z i ( n )∣ = l = 1 ∑ K κ l ( n ) w l p l ( n ) ,
where the second line groups the desks by cluster.
Activity driven by earlier updates. Lemma 4.2 links the size of the updates in a step to the activity of the same step. To close the loop, we now link the activity to the size of the updates in earlier steps. By (5) , each cluster averages the size of the updates over its horizon T k T_k T k , and we let
ω k ( s ) = 1 T k ( 1 − 1 T k ) s − 1 , s ∈ N ∖ { 0 } , \omega_k(s) = \frac{1}{T_k}\left(1-\frac{1}{T_k}\right)^{s-1}, \qquad s \in\mathbb{N} \setminus\{0\}, ω k ( s ) = T k 1 ( 1 − T k 1 ) s − 1 , s ∈ N ∖ { 0 } ,
denote the weight that this average gives to the size of the updates from s s s steps ago. As T k > 1 T_k > 1 T k > 1 , these weights are positive, sum to one over s s s and decay over about T k T_k T k steps.
Proposition 4.3 (Activity driven by earlier updates). Under Assumption 3.3(a), for every step n n n ,
A n = p 0 + β ∑ m = 0 n − 1 ( ∑ k = 1 K w k ω k ( n − m ) ) I m . A_n = p_0 + \beta\sum_{m=0}^{n-1}\left(\sum_{k=1}^{K} w_k \omega_k(n-m)\right) I_m. A n = p 0 + β m = 0 ∑ n − 1 ( k = 1 ∑ K w k ω k ( n − m ) ) I m .
The size of the updates in every step thus raises the activity of later steps through the memories of all clusters, each weighted by the volume of its cluster.
Proof. By Assumption 3.3(a) we have p 0 + β U k ( n ) = p 0 + β u ˉ k ( n ) ≤ p ˉ < 1 p_0 + \beta U_k(n) = p_0 + \beta\bar{u}_k(n) \le\bar{p} < 1 p 0 + β U k ( n ) = p 0 + β u ˉ k ( n ) ≤ p ˉ < 1 , such that (8) reads p k ( n ) = p 0 + β U k ( n ) p_k(n) = p_0 + \beta U_k(n) p k ( n ) = p 0 + β U k ( n ) . By (5) the estimates evolve as U k ( n ) = ( 1 − 1 / T k ) U k ( n − 1 ) + I n − 1 / T k U_k(n) = (1 - 1/T_k) U_k(n - 1) + I_{n-1}/T_k U k ( n ) = ( 1 − 1/ T k ) U k ( n − 1 ) + I n − 1 / T k from U k = I − 1 = 0 U_k = I_{-1} = 0 U k = I − 1 = 0 in § 2.6 . We show by induction that
U k ( n ) = ∑ m = 0 n − 1 ω k ( n − m ) I m . U_k(n) = \sum_{m=0}^{n-1} \omega_k(n-m) I_m. U k ( n ) = m = 0 ∑ n − 1 ω k ( n − m ) I m .
For n = 0 n=0 n = 0 both sides vanish. If the identity holds for n − 1 n-1 n − 1 , then using ( 1 − 1 / T k ) ω k ( s ) = ω k ( s + 1 ) (1-1/T_k)\omega_k(s)=\omega_k(s+1) ( 1 − 1/ T k ) ω k ( s ) = ω k ( s + 1 ) and ω k ( 1 ) = 1 / T k \omega_k(1)=1/T_k ω k ( 1 ) = 1/ T k , we obtain
U k ( n ) = ( 1 − 1 T k ) ∑ m = 0 n − 2 ω k ( n − 1 − m ) I m + I n − 1 T k = ∑ m = 0 n − 2 ω k ( n − m ) I m + ω k ( 1 ) I n − 1 . \begin{aligned}
U_k(n)=\left(1-\frac{1}{T_k}\right)\sum_{m=0}^{n-2}\omega_k(n-1-m)I_m+\frac{I_{n-1}}{T_k} \\
&=\sum_{m=0}^{n-2}\omega_k(n-m)I_m+\omega_k(1)I_{n-1}.
\end{aligned} U k ( n ) = ( 1 − T k 1 ) m = 0 ∑ n − 2 ω k ( n − 1 − m ) I m + T k I n − 1 = m = 0 ∑ n − 2 ω k ( n − m ) I m + ω k ( 1 ) I n − 1 .
Multiplying by β w k \beta w_k β w k , summing over k k k and using ∑ k w k = 1 \sum_k w_k=1 ∑ k w k = 1 from (4) , we obtain the claim. □
Two time scales. Lemma 4.2 and Proposition 4.3 together make the activity depend on its own past, except that κ l ( n ) \kappa_l(n) κ l ( n ) depends on the book. A larger size of the updates moves the mid price further. By (3) , every order then shifts from x i , j x_{i,j} x i , j to x i , j − σ i θ n + 1 > 0 x_{i,j}-\sigma_i\theta_{n+1}>0 x i , j − σ i θ n + 1 > 0 , and κ l \kappa_l κ l changes with the book. We now compare how fast the book recovers from such a shift with how long the shift stays in the moving averages (5) .
Lemma 4.4 (Two time scales). Let Assumptions 3.2 and 3.3 hold with γ < τ 1 \gamma<\tau_1 γ < τ 1 , and let o i , j o_{i,j} o i , j be an order of a desk of cluster k k k . Then
(i) its gap to the target after step n n n satisfies
E [ x i , j n e w − x i , j ∗ ∣ F n ] = ( 1 − γ p k ( n ) ) ( x i , j − x i , j ∗ ) + p k ( n ) σ i Π P k ; \mathbb{E}\left[x_{i,j}^{\mathrm{new}}-x_{i,j}^{*}\mid\mathcal{F}_n\right]=(1-\gamma p_k(n))(x_{i,j}-x_{i,j}^{*})+p_k(n)\sigma_i\Pi P_k; E [ x i , j new − x i , j ∗ ∣ F n ] = ( 1 − γ p k ( n )) ( x i , j − x i , j ∗ ) + p k ( n ) σ i Π P k ; (ii) for every s ∈ N ∖ { 0 } s\in\mathbb{N}\setminus\{0\} s ∈ N ∖ { 0 } , with t = ( s − 1 ) γ t=(s-1)\gamma t = ( s − 1 ) γ ,
γ τ k e − t / ( τ k − γ ) ≤ ω k ( s ) ≤ γ τ k e − t / τ k . \frac{\gamma}{\tau_k}e^{-t/(\tau_k-\gamma)}\leq\omega_k(s)\leq\frac{\gamma}{\tau_k}e^{-t/\tau_k}. τ k γ e − t / ( τ k − γ ) ≤ ω k ( s ) ≤ τ k γ e − t / τ k .
In expectation, each step thus shrinks the gap of an order to its target by the factor 1 − γ p k 1-\gamma p_k 1 − γ p k , apart from the move with the trend P k P_k P k . Over the time t t t , that is, over t / γ t/\gamma t / γ steps, the gap therefore shrinks by ( 1 − γ p k ) t / γ ≤ e − p k t (1-\gamma p_k)^{t/\gamma}\leq e^{-p_k t} ( 1 − γ p k ) t / γ ≤ e − p k t . By (ii), the memory of cluster k k k instead decays as e − t / τ k e^{-t/\tau_k} e − t / τ k .
Proof. For (i), by Assumption 3.3(b) we have x i , j − ξ i ( n ) γ ∂ j L i ( x i ) > 0 x_{i,j}-\xi_i(n)\gamma\partial_jL_i(x_i)>0 x i , j − ξ i ( n ) γ ∂ j L i ( x i ) > 0 , such that (9) reads x i , j n e w = x i , j − ξ i ( n ) γ ∂ j L i ( x i ) x_{i,j}^{\mathrm{new}}=x_{i,j}-\xi_i(n)\gamma\partial_jL_i(x_i) x i , j new = x i , j − ξ i ( n ) γ ∂ j L i ( x i ) . By (6) we have γ ∂ j L i ( x i ) = γ ( x i , j − x i , j ∗ ) − σ i Π P k \gamma\partial_jL_i(x_i)=\gamma(x_{i,j}-x_{i,j}^{*})-\sigma_i\Pi P_k γ ∂ j L i ( x i ) = γ ( x i , j − x i , j ∗ ) − σ i Π P k , and hence
x i , j n e w − x i , j ∗ = x i , j − x i , j ∗ − ξ i ( n ) ( γ ( x i , j − x i , j ∗ ) − σ i Π P k ) = ( 1 − γ ξ i ( n ) ) ( x i , j − x i , j ∗ ) + ξ i ( n ) σ i Π P k . \begin{aligned}
x_{i,j}^{\mathrm{new}}-x_{i,j}^{*}=x_{i,j}-x_{i,j}^{*}-\xi_i(n)\left(\gamma(x_{i,j}-x_{i,j}^{*})-\sigma_i\Pi P_k\right) \\
&=(1-\gamma\xi_i(n))(x_{i,j}-x_{i,j}^{*})+\xi_i(n)\sigma_i\Pi P_k.
\end{aligned} x i , j new − x i , j ∗ = x i , j − x i , j ∗ − ξ i ( n ) ( γ ( x i , j − x i , j ∗ ) − σ i Π P k ) = ( 1 − γ ξ i ( n )) ( x i , j − x i , j ∗ ) + ξ i ( n ) σ i Π P k .
As x i , j x_{i,j} x i , j and P k P_k P k are part of X n X_n X n , we obtain (i) by taking the conditional expectation with E [ ξ i ( n ) ∣ F n ] = p k ( n ) \mathbb{E}[\xi_i(n)\mid\mathcal{F}_n]=p_k(n) E [ ξ i ( n ) ∣ F n ] = p k ( n ) . For (ii), T k = τ k / γ T_k=\tau_k/\gamma T k = τ k / γ by Assumption 3.2, such that ω k ( s ) = ( γ / τ k ) ( 1 − γ / τ k ) s − 1 \omega_k(s)=(\gamma/\tau_k)(1-\gamma/\tau_k)^{s-1} ω k ( s ) = ( γ / τ k ) ( 1 − γ / τ k ) s − 1 . For 0 ≤ y < 1 0\leq y<1 0 ≤ y < 1 we have − y / ( 1 − y ) ≤ ln ( 1 − y ) ≤ − y -y/(1-y)\leq\ln(1-y)\leq-y − y / ( 1 − y ) ≤ ln ( 1 − y ) ≤ − y . With y = γ / τ k y=\gamma/\tau_k y = γ / τ k , multiplying by s − 1 s-1 s − 1 and taking exponentials, we obtain
e − ( s − 1 ) γ / ( τ k − γ ) ≤ ( 1 − γ τ k ) s − 1 ≤ e − ( s − 1 ) γ / τ k , e^{-(s-1)\gamma/(\tau_k-\gamma)}\leq\left(1-\frac{\gamma}{\tau_k}\right)^{s-1}\leq e^{-(s-1)\gamma/\tau_k}, e − ( s − 1 ) γ / ( τ k − γ ) ≤ ( 1 − τ k γ ) s − 1 ≤ e − ( s − 1 ) γ / τ k ,
which is (ii). □
The book thus recovers from a shift within a time of order 1 / p k 1/p_k 1/ p k , while the moving average (5) of cluster k k k decays over a time of order τ k \tau_k τ k . Whenever τ k p k ≫ 1 \tau_kp_k\gg1 τ k p k ≫ 1 , the book recovers its shape long before the moving averages decay. We therefore replace κ l ( n ) \kappa_l(n) κ l ( n ) by its average κ l \kappa_l κ l along the path, which we verify in § 6 . With these averages we write
ϕ ( s ) = β ∑ l = 1 K κ l w l ω l ( s ) , κ ˉ = ∑ l = 1 K κ l w l , (15) \phi(s) = \beta\sum_{l=1}^{K} \kappa_l w_l \omega_l(s), \qquad\bar{\kappa} = \sum_{l=1}^{K} \kappa_l w_l, \tag*{(15)} ϕ ( s ) = β l = 1 ∑ K κ l w l ω l ( s ) , κ ˉ = l = 1 ∑ K κ l w l , (15)
which combine the memories ω l \omega_l ω l , the volumes w l w_l w l and the sizes of the updates per unit activity κ l \kappa_l κ l of all clusters.
Proposition 4.5 (Self-excitation). Let the assumptions of Proposition 4.3 hold, and let κ l ( n ) = κ l \kappa_l(n) = \kappa_l κ l ( n ) = κ l for every cluster l l l and step n n n . Then the size of the updates (I n I_n I n ) is self-exciting in the sense of Definition 4.1 , with the base rate μ = p 0 κ ˉ \mu= p_0 \bar{\kappa} μ = p 0 κ ˉ and the kernel ϕ ( s ) = β ∑ l κ l w l ω l ( s ) \phi(s) = \beta\sum_l \kappa_l w_l \omega_l(s) ϕ ( s ) = β ∑ l κ l w l ω l ( s ) of (15) . In particular,
E [ I n ] = p 0 κ ˉ + ∑ m = 0 n − 1 ϕ ( n − m ) E [ I m ] . (16) \mathbb{E}[I_n] = p_0 \bar{\kappa} + \sum_{m=0}^{n-1} \phi(n-m) \mathbb{E}[I_m]. \tag*{(16)} E [ I n ] = p 0 κ ˉ + m = 0 ∑ n − 1 ϕ ( n − m ) E [ I m ] . (16)
Although no kernel enters the protocol, the desks thus make the size of the updates excite itself, with a kernel built from their memories, their volumes and the book.
Proof. The size of the updates I n I_n I n is nonnegative by (7) and F n + 1 \mathcal{F}_{n+1} F n + 1 -measurable. Combining Lemma 4.2 with the formula for U l ( n ) U_l(n) U l ( n ) from the proof of Proposition 4.3 , we obtain
E [ I n ∣ F n ] = ∑ l = 1 K κ l w l ( p 0 + β ∑ m = 0 n − 1 ω l ( n − m ) I m ) = p 0 κ ˉ + ∑ m = 0 n − 1 ( β ∑ l = 1 K κ l w l ω l ( n − m ) ) I m , \begin{aligned}
\mathbb{E}[I_n \mid\mathcal{F}_n] = \sum_{l=1}^{K} \kappa_l w_l \left(p_0 + \beta\sum_{m=0}^{n-1} \omega_l(n-m) I_m\right) \\
&= p_0 \bar{\kappa} + \sum_{m=0}^{n-1} \left(\beta\sum_{l=1}^{K} \kappa_l w_l \omega_l(n-m)\right) I_m,
\end{aligned} E [ I n ∣ F n ] = l = 1 ∑ K κ l w l ( p 0 + β m = 0 ∑ n − 1 ω l ( n − m ) I m ) = p 0 κ ˉ + m = 0 ∑ n − 1 ( β l = 1 ∑ K κ l w l ω l ( n − m ) ) I m ,
where the inner sum is ϕ ( n − m ) \phi(n-m) ϕ ( n − m ) by (15) . Taking expectations, we obtain (16) . □
Response to one extra update. We now follow a single extra update through the market. To this end we add E > 0 E > 0 E > 0 to I m 0 I_{m_0} I m 0 in (5) and leave everything else unchanged. The estimates of step m 0 + 1 m_0+1 m 0 + 1 then receive I m 0 + E I_{m_0}+E I m 0 + E , and the estimates of every later step n n n contain the additional term E ω l ( n − m 0 ) E\omega_l(n-m_0) E ω l ( n − m 0 ) . Let δ n \delta_n δ n denote the difference of E [ I n ] \mathbb{E}[I_n] E [ I n ] between the disturbed and the undisturbed market, with δ n = 0 \delta_n = 0 δ n = 0 for n ≤ m 0 n \le m_0 n ≤ m 0 . By (16) , we obtain for s ∈ N ∖ { 0 } s \in\mathbb{N} \setminus\{0\} s ∈ N ∖ { 0 }
δ m 0 + s = E ϕ ( s ) + ∑ r = 1 s − 1 ϕ ( s − r ) δ m 0 + r . \delta_{m_0+s} = E\phi(s) + \sum_{r=1}^{s-1} \phi(s-r)\delta_{m_0+r}. δ m 0 + s = E ϕ ( s ) + r = 1 ∑ s − 1 ϕ ( s − r ) δ m 0 + r .
Substituting this identity into its own right-hand side, we obtain
δ m 0 + s = E ϕ ( s ) + E ∑ r = 1 s − 1 ϕ ( s − r ) ϕ ( r ) + ∑ r = 1 s − 1 ∑ q = 1 r − 1 ϕ ( s − r ) ϕ ( r − q ) δ m 0 + q = E ϕ ∗ 1 ( s ) + E ϕ ∗ 2 ( s ) + E ϕ ∗ 3 ( s ) + ⋯ , \begin{aligned}
\delta_{m_0+s} = E\phi(s) + E\sum_{r=1}^{s-1}\phi(s-r)\phi(r) + \sum_{r=1}^{s-1}\sum_{q=1}^{r-1}\phi(s-r)\phi(r-q)\delta_{m_0+q} \\
&= E\phi^{*1}(s) + E\phi^{*2}(s) + E\phi^{*3}(s) + \cdots,
\end{aligned} δ m 0 + s = E ϕ ( s ) + E r = 1 ∑ s − 1 ϕ ( s − r ) ϕ ( r ) + r = 1 ∑ s − 1 q = 1 ∑ r − 1 ϕ ( s − r ) ϕ ( r − q ) δ m 0 + q = E ϕ ∗ 1 ( s ) + E ϕ ∗ 2 ( s ) + E ϕ ∗ 3 ( s ) + ⋯ ,
where ϕ ∗ 1 = ϕ \phi^{*1} = \phi ϕ ∗ 1 = ϕ and ϕ ∗ ( j + 1 ) ( s ) = ∑ r = 1 s − 1 ϕ ( s − r ) ϕ ∗ j ( r ) \phi^{*(j+1)}(s) = \sum_{r=1}^{s-1}\phi(s-r)\phi^{*j}(r) ϕ ∗ ( j + 1 ) ( s ) = ∑ r = 1 s − 1 ϕ ( s − r ) ϕ ∗ j ( r ) . The term E ϕ ∗ j ( s ) E\phi^{*j}(s) E ϕ ∗ j ( s ) is the contribution of E E E through j − 1 j-1 j − 1 intermediate updates, and it vanishes for j > s j>s j > s . We therefore define the total response
R ( s ) = ∑ j ∈ N ∖ { 0 } ϕ ∗ j ( s ) , s ∈ N ∖ { 0 } . R(s) = \sum_{j\in\mathbb{N}\setminus\{0\}} \phi^{*j}(s), \qquad s \in\mathbb{N}\setminus\{0\}. R ( s ) = j ∈ N ∖ { 0 } ∑ ϕ ∗ j ( s ) , s ∈ N ∖ { 0 } .
Corollary 4.6 (Response to one extra update). Under the assumptions of Proposition 4.5 , adding E E E to I m 0 I_{m_0} I m 0 in (5) leaves E [ I n ] \mathbb{E}[I_n] E [ I n ] unchanged for n ≤ m 0 n \le m_0 n ≤ m 0 and raises E [ I m 0 + s ] \mathbb{E}[I_{m_0+s}] E [ I m 0 + s ] by E R ( s ) E R(s) E R ( s ) for every s ∈ N ∖ { 0 } s \in\mathbb{N} \setminus\{0\} s ∈ N ∖ { 0 } .
The response to one extra update is thus fixed by ϕ \phi ϕ alone, which involves only the horizons, the volumes, β \beta β and the sizes of the updates per unit activity.
Proof. Up to step m 0 m_0 m 0 nothing changes. Splitting off the term j = 1 j=1 j = 1 in the definition of R R R , we obtain
R ( s ) = ϕ ( s ) + ∑ j ∈ N ∖ { 0 } ϕ ∗ ( j + 1 ) ( s ) = ϕ ( s ) + ∑ r = 1 s − 1 ϕ ( s − r ) ∑ j ∈ N ∖ { 0 } ϕ ∗ j ( r ) = ϕ ( s ) + ∑ r = 1 s − 1 ϕ ( s − r ) R ( r ) . \begin{aligned}
R(s) = \phi(s) + \sum_{j\in\mathbb{N}\setminus\{0\}} \phi^{*(j+1)}(s) = \phi(s) + \sum_{r=1}^{s-1} \phi(s-r) \sum_{j\in\mathbb{N}\setminus\{0\}} \phi^{*j}(r) \\
&= \phi(s) + \sum_{r=1}^{s-1} \phi(s-r)R(r).
\end{aligned} R ( s ) = ϕ ( s ) + j ∈ N ∖ { 0 } ∑ ϕ ∗ ( j + 1 ) ( s ) = ϕ ( s ) + r = 1 ∑ s − 1 ϕ ( s − r ) j ∈ N ∖ { 0 } ∑ ϕ ∗ j ( r ) = ϕ ( s ) + r = 1 ∑ s − 1 ϕ ( s − r ) R ( r ) .
The difference δ m 0 + s \delta_{m_0+s} δ m 0 + s and E R ( s ) E R(s) E R ( s ) thus satisfy the same recursion. For s = 1 s=1 s = 1 both equal E ϕ ( 1 ) E\phi(1) E ϕ ( 1 ) , and by induction on s s s they agree for every s ∈ N ∖ { 0 } s \in\mathbb{N}\setminus\{0\} s ∈ N ∖ { 0 } . □
The desks thus produce a self-exciting order flow whose kernel we know explicitly. We next compute ∥ ϕ ∥ \|\phi\| ∥ ϕ ∥ for our market, which decides whether the activity stays finite.
Norm of the kernel and mean activity
By Definition 4.1 , the size of the updates stays finite on average only if ∥ ϕ ∥ < 1 \|\phi\| < 1 ∥ ϕ ∥ < 1 . By [41 ] , roughness in turn requires self-excitation close to this threshold, with ∥ ϕ ∥ \|\phi\| ∥ ϕ ∥ close to one. Whether our market can come close to the threshold depends on what sets ∥ ϕ ∥ \|\phi\| ∥ ϕ ∥ . We now compute ∥ ϕ ∥ \|\phi\| ∥ ϕ ∥ in terms of the desks and the book, show how it fixes the mean activity, and write the kernel as a function of time.
Norm of the kernel. Each memory ω l \omega_l ω l sums to one over s s s , such that summing (15) over s s s leaves only the sensitivity β \beta β and the sizes of the updates per unit activity.
Proposition 4.7 (Norm of the kernel). Under the assumptions of Proposition 4.5 ,
∥ ϕ ∥ = β κ ˉ = β ∑ l = 1 K κ l w l . \|\phi\| = \beta\bar{\kappa} = \beta\sum_{l=1}^{K} \kappa_l w_l. ∥ ϕ ∥ = β κ ˉ = β l = 1 ∑ K κ l w l . Moreover, a unit of activity of cluster l l l in step m m m raises the expected activity of all later steps together by β κ l \beta\kappa_l β κ l .
A unit of activity of cluster l l l thus sets off on average β κ l \beta\kappa_l β κ l units of activity in later steps. Since κ l \kappa_l κ l is large when the orders of cluster l l l sit near the mid price, a book with more volume near the mid price excites itself more strongly.
Proof. As ∑ s ∈ N ∖ { 0 } ( 1 − 1 / T l ) s − 1 = T l \sum_{s\in\mathbb{N}\setminus\{0\}}(1-1/T_l)^{s-1}=T_l ∑ s ∈ N ∖ { 0 } ( 1 − 1/ T l ) s − 1 = T l , each memory satisfies ∑ s ω l ( s ) = 1 \sum_s\omega_l(s)=1 ∑ s ω l ( s ) = 1 . Summing (15) over s s s , we obtain
∥ ϕ ∥ = β ∑ l = 1 K κ l w l ∑ s ∈ N ∖ { 0 } ω l ( s ) = β ∑ l = 1 K κ l w l = β κ ˉ . \|\phi\| = \beta\sum_{l=1}^{K} \kappa_l w_l \sum_{s\in\mathbb{N}\setminus\{0\}} \omega_l(s) = \beta\sum_{l=1}^{K} \kappa_l w_l = \beta\bar{\kappa}. ∥ ϕ ∥ = β l = 1 ∑ K κ l w l s ∈ N ∖ { 0 } ∑ ω l ( s ) = β l = 1 ∑ K κ l w l = β κ ˉ .
For the second claim, by Lemma 4.2 a unit of activity of cluster l l l in step m m m adds κ l \kappa_l κ l to the expected size of the updates in step m m m . By Proposition 4.3 this addition raises the activity of step n > m n > m n > m by β κ l ∑ k w k ω k ( n − m ) \beta\kappa_l \sum_{k} w_k\omega_k(n-m) β κ l ∑ k w k ω k ( n − m ) . Summing over n > m n > m n > m and using ∑ k w k = 1 \sum_k w_k = 1 ∑ k w k = 1 from (4) , we obtain
∑ n > m β κ l ∑ k = 1 K w k ω k ( n − m ) = β κ l ∑ k = 1 K w k ∑ s ∈ N ∖ { 0 } ω k ( s ) = β κ l . \sum_{n>m}\beta\kappa_l\sum_{k=1}^{K}w_k\omega_k(n-m)=\beta\kappa_l\sum_{k=1}^{K}w_k\sum_{s\in\mathbb{N}\setminus\{0\}}\omega_k(s)=\beta\kappa_l. n > m ∑ β κ l k = 1 ∑ K w k ω k ( n − m ) = β κ l k = 1 ∑ K w k s ∈ N ∖ { 0 } ∑ ω k ( s ) = β κ l .
Mean activity. By Definition 4.1, the threshold ∥ ϕ ∥ = 1 \lVert\phi\rVert= 1 ∥ ϕ ∥ = 1 decides whether the size of the updates stays finite on average. As the base rate p 0 κ ˉ p_0\bar{\kappa} p 0 κ ˉ of our market does not change over time, we can solve (16) exactly and follow the activity on both sides of the threshold.
Proposition 4.8 (Mean activity). Let the assumptions of Proposition 4.5 hold with κ ˉ > 0 \bar{\kappa} > 0 κ ˉ > 0 . Then for every step n n n ,
E [ I n ] = p 0 κ ˉ ( 1 + ∑ s = 1 n R ( s ) ) . \mathbb{E}[I_n]=p_0\bar{\kappa}\left(1+\sum_{s=1}^{n}R(s)\right). E [ I n ] = p 0 κ ˉ ( 1 + s = 1 ∑ n R ( s ) ) . (i) If ∥ ϕ ∥ < 1 \lVert\phi\rVert< 1 ∥ ϕ ∥ < 1 , then ∑ s ∈ N ∖ { 0 } R ( s ) = ∥ ϕ ∥ / ( 1 − ∥ ϕ ∥ ) \sum_{s\in\mathbb{N}\setminus\{0\}}R(s)=\lVert\phi\rVert/(1-\lVert\phi\rVert) ∑ s ∈ N ∖ { 0 } R ( s ) = ∥ ϕ ∥ / ( 1 − ∥ ϕ ∥) , and as n → ∞ n\to\infty n → ∞
E [ I n ] ↑ p 0 κ ˉ 1 − ∥ ϕ ∥ , E [ p k ( n ) ] → p 0 1 − ∥ ϕ ∥ for every cluster k , E [ A n ] → p 0 1 − ∥ ϕ ∥ . \mathbb{E}[I_n]\uparrow\frac{p_0\bar{\kappa}}{1-\lVert\phi\rVert},\qquad\mathbb{E}[p_k(n)]\to\frac{p_0}{1-\lVert\phi\rVert}\quad\text{for every cluster }k,\qquad\mathbb{E}[A_n]\to\frac{p_0}{1-\lVert\phi\rVert}. E [ I n ] ↑ 1 − ∥ ϕ ∥ p 0 κ ˉ , E [ p k ( n )] → 1 − ∥ ϕ ∥ p 0 for every cluster k , E [ A n ] → 1 − ∥ ϕ ∥ p 0 . In particular, Assumption 3.3(a) implies ∥ ϕ ∥ ≤ 1 − p 0 / p ˉ \lVert\phi\rVert\leq1-p_0/\bar{p} ∥ ϕ ∥ ≤ 1 − p 0 / p ˉ .
(ii) If ∥ ϕ ∥ ≥ 1 \lVert\phi\rVert\geq1 ∥ ϕ ∥ ≥ 1 , then there exist a cluster k ∈ { 1 , … , K } k\in\{1,\ldots,K\} k ∈ { 1 , … , K } and a step n ∈ N n\in\mathbb{N} n ∈ N with
P ( p 0 + β U k ( n ) > p ˉ ) > 0. \mathbb{P}(p_0+\beta U_k(n)>\bar{p})>0. P ( p 0 + β U k ( n ) > p ˉ ) > 0.
Below the threshold, the desks thus multiply the activity p 0 p_0 p 0 , which comes from information outside the desks, by the factor 1 / ( 1 − ∥ ϕ ∥ ) 1/(1-\lVert\phi\rVert) 1/ ( 1 − ∥ ϕ ∥) . As ∥ ϕ ∥ \lVert\phi\rVert ∥ ϕ ∥ approaches one, almost all activity is set off by earlier actions of the desks, which by (i) requires p 0 p_0 p 0 to be small. However, above this threshold, the probabilities of acting grow until they exceed p ˉ \bar{p} p ˉ , which leads to the failure of Assumption 3.3(a).
Proof. We first show the formula for E [ I n ] \mathbb{E}[I_n] E [ I n ] by induction on n n n . For n = 0 n=0 n = 0 we obtain E [ I 0 ] = p 0 κ ˉ \mathbb{E}[I_0]=p_0\bar{\kappa} E [ I 0 ] = p 0 κ ˉ from (16) . If the formula holds below n n n , then substituting it into (16) with u = n − m u=n-m u = n − m , we obtain
E [ I n ] = p 0 κ ˉ + ∑ m = 0 n − 1 ϕ ( n − m ) p 0 κ ˉ ( 1 + ∑ r = 1 m R ( r ) ) = p 0 κ ˉ ( 1 + ∑ u = 1 n ϕ ( u ) + ∑ r = 1 n − 1 R ( r ) ∑ u = 1 n − r ϕ ( u ) ) = p 0 κ ˉ ( 1 + ∑ s = 1 n ( ϕ ( s ) + ∑ r = 1 s − 1 ϕ ( s − r ) R ( r ) ) ) = p 0 κ ˉ ( 1 + ∑ s = 1 n R ( s ) ) . \begin{aligned}
\mathbb{E}[I_n]=p_0\bar{\kappa}+\sum_{m=0}^{n-1}\phi(n-m)p_0\bar{\kappa}\left(1+\sum_{r=1}^{m}R(r)\right) \\
&=p_0\bar{\kappa}\left(1+\sum_{u=1}^{n}\phi(u)+\sum_{r=1}^{n-1}R(r)\sum_{u=1}^{n-r}\phi(u)\right) \\
&=p_0\bar{\kappa}\left(1+\sum_{s=1}^{n}\left(\phi(s)+\sum_{r=1}^{s-1}\phi(s-r)R(r)\right)\right)=p_0\bar{\kappa}\left(1+\sum_{s=1}^{n}R(s)\right).
\end{aligned} E [ I n ] = p 0 κ ˉ + m = 0 ∑ n − 1 ϕ ( n − m ) p 0 κ ˉ ( 1 + r = 1 ∑ m R ( r ) ) = p 0 κ ˉ ( 1 + u = 1 ∑ n ϕ ( u ) + r = 1 ∑ n − 1 R ( r ) u = 1 ∑ n − r ϕ ( u ) ) = p 0 κ ˉ ( 1 + s = 1 ∑ n ( ϕ ( s ) + r = 1 ∑ s − 1 ϕ ( s − r ) R ( r ) ) ) = p 0 κ ˉ ( 1 + s = 1 ∑ n R ( s ) ) .
where the third line collects the terms with r + u = s r+u=s r + u = s and the last equality is the recursion R ( s ) = ϕ ( s ) + ∑ r = 1 s − 1 ϕ ( s − r ) R ( r ) R(s)=\phi(s)+\sum_{r=1}^{s-1}\phi(s-r)R(r) R ( s ) = ϕ ( s ) + ∑ r = 1 s − 1 ϕ ( s − r ) R ( r ) , which we showed in the proof of Corollary 4.6. Next, for functions ϕ , ψ : N ∖ { 0 } → [ 0 , ∞ ) \phi,\psi:\mathbb{N}\setminus\{0\}\to[0,\infty) ϕ , ψ : N ∖ { 0 } → [ 0 , ∞ ) we have ∥ ϕ ∗ ψ ∥ = ∥ ϕ ∥ ∥ ψ ∥ \lVert\phi* \psi\rVert=\lVert\phi\rVert\lVert\psi\rVert ∥ ϕ ∗ ψ ∥ = ∥ ϕ ∥ ∥ ψ ∥ , as both sides sum the same nonnegative terms. Hence ∥ ϕ ∗ j ∥ = ∥ ϕ ∥ j \lVert\phi^{*j}\rVert=\lVert\phi\rVert^j ∥ ϕ ∗ j ∥ = ∥ ϕ ∥ j , and
∑ s ∈ N ∖ { 0 } R ( s ) = ∑ j ∈ N ∖ { 0 } ∥ ϕ ∥ j = { ∥ ϕ ∥ / ( 1 − ∥ ϕ ∥ ) , ∥ ϕ ∥ < 1 , ∞ , ∥ ϕ ∥ ≥ 1. \sum_{s\in\mathbb{N}\setminus\{0\}}R(s)=\sum_{j\in\mathbb{N}\setminus\{0\}}\lVert\phi\rVert^j=
\begin{cases}
\lVert\phi\rVert/(1-\lVert\phi\rVert), & \lVert\phi\rVert<1,\\
\infty, & \lVert\phi\rVert\geq1.
\end{cases} s ∈ N ∖ { 0 } ∑ R ( s ) = j ∈ N ∖ { 0 } ∑ ∥ ϕ ∥ j = { ∥ ϕ ∥ / ( 1 − ∥ ϕ ∥) , ∞ , ∥ ϕ ∥ < 1 , ∥ ϕ ∥ ≥ 1.
For (i), the partial sums of R ≥ 0 R\geq0 R ≥ 0 increase to ∥ ϕ ∥ / ( 1 − ∥ ϕ ∥ ) \lVert\phi\rVert/(1-\lVert\phi\rVert) ∥ ϕ ∥ / ( 1 − ∥ ϕ ∥) , such that E [ I n ] \mathbb{E}[I_n] E [ I n ] increases to I ∗ = p 0 κ ˉ / ( 1 − ∥ ϕ ∥ ) I^*=p_0\bar{\kappa}/(1-\lVert\phi\rVert) I ∗ = p 0 κ ˉ / ( 1 − ∥ ϕ ∥) . By the proof of Proposition 4.3, E [ p k ( n ) ] = p 0 + β ∑ s = 1 n ω k ( s ) E [ I n − s ] \mathbb{E}[p_k(n)]=p_0+\beta\sum_{s=1}^{n}\omega_k(s)\mathbb{E}[I_{n-s}] E [ p k ( n )] = p 0 + β ∑ s = 1 n ω k ( s ) E [ I n − s ] . Each term converges to β ω k ( s ) I ∗ \beta\omega_k(s) I^* β ω k ( s ) I ∗ and is bounded by it, while ∑ s ω k ( s ) = 1 \sum_s \omega_k(s)=1 ∑ s ω k ( s ) = 1 . By dominated convergence and Proposition 4.7 , we therefore obtain
E [ p k ( n ) ] → p 0 + β I ∗ = p 0 ( 1 + β κ ˉ 1 − ∥ ϕ ∥ ) = p 0 1 − ∥ ϕ ∥ , \mathbb{E}[p_k(n)] \to p_0+\beta I^*=p_0\left(1+\frac{\beta\bar{\kappa}}{1-\lVert\phi\rVert}\right)=\frac{p_0}{1-\lVert\phi\rVert}, E [ p k ( n )] → p 0 + β I ∗ = p 0 ( 1 + 1 − ∥ ϕ ∥ β κ ˉ ) = 1 − ∥ ϕ ∥ p 0 ,
and E [ A n ] = ∑ k w k E [ p k ( n ) ] \mathbb{E}[A_n]=\sum_k w_k\mathbb{E}[p_k(n)] E [ A n ] = ∑ k w k E [ p k ( n )] has the same limit. By Assumption 3.3 (a) we have p k ( n ) ≤ p ˉ p_k(n)\leq\bar{p} p k ( n ) ≤ p ˉ in every step, such that the limit satisfies p 0 / ( 1 − ∥ ϕ ∥ ) ≤ p ˉ p_0/(1-\lVert\phi\rVert)\leq\bar{p} p 0 / ( 1 − ∥ ϕ ∥) ≤ p ˉ , that is, ∥ ϕ ∥ ≤ 1 − p 0 / p ˉ \lVert\phi\rVert\leq1-p_0/\bar{p} ∥ ϕ ∥ ≤ 1 − p 0 / p ˉ . For (ii), suppose that p 0 + β U k ( n ) ≤ p ˉ p_0+\beta U_k(n)\leq\bar{p} p 0 + β U k ( n ) ≤ p ˉ almost surely for every cluster k k k and step n n n , such that Assumption 3.3 (a) holds in every step. By Lemma 4.2 and Assumption 3.3 (a), we then obtain E [ I n ] = ∑ l κ l w l E [ p l ( n ) ] ≤ p ˉ κ ˉ \mathbb{E}[I_n]=\sum_l\kappa_lw_l\mathbb{E}[p_l(n)]\leq\bar{p}\bar{\kappa} E [ I n ] = ∑ l κ l w l E [ p l ( n )] ≤ p ˉ κ ˉ for every n n n . By the formula for E [ I n ] \mathbb{E}[I_n] E [ I n ] and ∑ s R ( s ) = ∞ \sum_s R(s)=\infty ∑ s R ( s ) = ∞ , however, E [ I n ] → ∞ \mathbb{E}[I_n]\to\infty E [ I n ] → ∞ , which is a contradiction. □
The kernel as a function of time. The kernel ϕ ( s ) \phi(s) ϕ ( s ) measures the lag s s s in steps. Note, however, that in the continuum order book (Proposition 3.10 ) each step lasts only the time γ \gamma γ , while by Assumption 3.2 the moving average (5) of cluster l l l extends over T l = τ l / γ T_l=\tau_l/\gamma T l = τ l / γ steps. Therefore, as γ → 0 + \gamma\to0^+ γ → 0 + , each value ϕ ( s ) \phi(s) ϕ ( s ) vanishes, although each moving average still covers the fixed time T l γ = τ l T_l\gamma=\tau_l T l γ = τ l . We thus write the kernel as a function of the elapsed time t = ( s − 1 ) γ t=(s-1)\gamma t = ( s − 1 ) γ . Expanding the exponentials in Δ Z i \Delta Z_i Δ Z i to first order in γ \gamma γ , an acting desk changes its volume near the mid price by ∣ Δ Z i ∣ = γ ∣ a i ∣ + O ( γ 2 ) \lvert\Delta Z_i\rvert=\gamma\lvert a_i\rvert+O(\gamma^2) ∣ Δ Z i ∣ = γ ∣ a i ∣ + O ( γ 2 ) with a i = − ( σ i / ℓ ) ∑ j = 1 m i e − σ i y i , j / ℓ b i , j a_i=-(\sigma_i/\ell)\sum_{j=1}^{m_i}e^{-\sigma_i y_{i,j}/\ell}b_{i,j} a i = − ( σ i / ℓ ) ∑ j = 1 m i e − σ i y i , j / ℓ b i , j , while c = γ c ˉ c=\gamma\bar{c} c = γ c ˉ by Assumption 3.2 . The sizes of the updates per unit activity thus satisfy
κ l = γ 2 c ˉ w l ∑ i : k ( i ) = l m i ∣ a i ∣ + O ( γ 3 ) , \kappa_l=\gamma^2\frac{\bar{c}}{w_l}\sum_{i:k(i)=l}m_i\lvert a_i\rvert+O(\gamma^3), κ l = γ 2 w l c ˉ i : k ( i ) = l ∑ m i ∣ a i ∣ + O ( γ 3 ) ,
such that ∥ ϕ ∥ = β κ ˉ \lVert\phi\rVert=\beta\bar{\kappa} ∥ ϕ ∥ = β κ ˉ with β = β ˉ / γ 2 \beta=\bar{\beta}/\gamma^2 β = β ˉ / γ 2 stays of order one as γ → 0 + \gamma\to0^+ γ → 0 + . Let
ϕ ˉ ( t ) = ∥ ϕ ∥ ∑ l = 1 K ϖ l e − t / τ l τ l , ϖ l = κ l w l κ ˉ , t ≥ 0 , \bar{\phi}(t)=\lVert\phi\rVert\sum_{l=1}^{K}\varpi_l\frac{e^{-t/\tau_l}}{\tau_l},\qquad\varpi_l=\frac{\kappa_lw_l}{\bar{\kappa}},\qquad t\geq0, ϕ ˉ ( t ) = ∥ ϕ ∥ l = 1 ∑ K ϖ l τ l e − t / τ l , ϖ l = κ ˉ κ l w l , t ≥ 0 ,
denote the kernel as a function of time, where the weights ϖ l \varpi_l ϖ l sum to one.
Proposition 4.9 (The kernel as a function of time). Under the assumptions of Proposition 4.5 and Assumption 3.2 with γ < τ 1 \gamma<\tau_1 γ < τ 1 , for every s ∈ N ∖ { 0 } s\in\mathbb{N}\setminus\{0\} s ∈ N ∖ { 0 } and t = ( s − 1 ) γ t=(s-1)\gamma t = ( s − 1 ) γ ,
γ ∥ ϕ ∥ ∑ l = 1 K ϖ l e − t / ( τ l − γ ) τ l ≤ ϕ ( s ) ≤ γ ϕ ˉ ( t ) , ∫ 0 ∞ ϕ ˉ ( t ) d t = ∥ ϕ ∥ . \gamma\lVert\phi\rVert\sum_{l=1}^{K}\varpi_l\frac{e^{-t/(\tau_l-\gamma)}}{\tau_l}\leq\phi(s)\leq\gamma\bar{\phi}(t),\qquad\int_0^\infty\bar{\phi}(t)\,\mathrm{d}t=\lVert\phi\rVert. γ ∥ ϕ ∥ l = 1 ∑ K ϖ l τ l e − t / ( τ l − γ ) ≤ ϕ ( s ) ≤ γ ϕ ˉ ( t ) , ∫ 0 ∞ ϕ ˉ ( t ) d t = ∥ ϕ ∥ .
As a function of time, the kernel is thus a mixture of exponentials, one for each horizon τ l \tau_l τ l , where cluster l l l enters with the weight ϖ l = κ l w l / κ ˉ \varpi_l=\kappa_lw_l/\bar{\kappa} ϖ l = κ l w l / κ ˉ . Its total weight ∥ ϕ ∥ \lVert\phi\rVert ∥ ϕ ∥ does not depend on γ \gamma γ and thus survives the limit γ → 0 + \gamma\to0^+ γ → 0 + .
Proof. By Proposition 4.7 we have β κ l w l = ∥ ϕ ∥ ϖ l \beta\kappa_lw_l=\lVert\phi\rVert\varpi_l β κ l w l = ∥ ϕ ∥ ϖ l . Multiplying Lemma 4.4 (ii) by β κ l w l \beta\kappa_lw_l β κ l w l and summing over l l l , we obtain the two bounds from (15) . Lastly, as ∫ 0 ∞ e − t / τ l d t / τ l = 1 \int_0^\infty e^{-t/\tau_l}\,\mathrm{d}t/\tau_l=1 ∫ 0 ∞ e − t / τ l d t / τ l = 1 and ∑ l ϖ l = 1 \sum_l\varpi_l=1 ∑ l ϖ l = 1 , we obtain ∫ 0 ∞ ϕ ˉ ( t ) d t = ∥ ϕ ∥ \int_0^\infty\bar{\phi}(t)\,\mathrm{d}t=\lVert\phi\rVert ∫ 0 ∞ ϕ ˉ ( t ) d t = ∥ ϕ ∥ . □
We next ask how this mixture fades when the horizons of the desks spread over many scales.
A power-law kernel from many horizons
By Proposition 4.9 , the kernel of our market is a mixture of exponentials, one for each horizon τ l \tau_l τ l ,
ϕ ˉ ( t ) = ∥ ϕ ∥ ∑ l = 1 K ϖ l e − t / τ l τ l , t ≥ 0. \bar{\phi}(t) = \lVert\phi\rVert\sum_{l=1}^{K}\varpi_l\frac{e^{-t/\tau_l}}{\tau_l}, \qquad t \ge0. ϕ ˉ ( t ) = ∥ ϕ ∥ l = 1 ∑ K ϖ l τ l e − t / τ l , t ≥ 0.
For a single horizon, K = 1 K=1 K = 1 , the weight of the kernel beyond the time x x x decays exponentially,
∫ x ∞ ϕ ˉ ( t ) d t = ∥ ϕ ∥ e − x / τ 1 , x ≥ 0. \int_x^\infty\bar{\phi}(t)\,\mathrm{d}t = \lVert\phi\rVert e^{-x/\tau_1}, \qquad x \ge0. ∫ x ∞ ϕ ˉ ( t ) d t = ∥ ϕ ∥ e − x / τ 1 , x ≥ 0.
The self-excitation of the size of the updates therefore extends only over times of order τ 1 \tau_1 τ 1 . In this subsection we show that horizons on many scales instead make ϕ ˉ \bar{\phi} ϕ ˉ decay as the power t − 1 − α t^{-1-\alpha} t − 1 − α with the exponent α \alpha α of the volumes (4) .
The horizons of the clusters, however, spread over many scales, as by § 2.3 and Assumption 3.2
τ k = γ T k = τ 1 Ω k − 1 , k ∈ { 1 , … , K } . \tau_k = \gamma T_k = \tau_1 \Omega^{k-1}, \qquad k \in\{1,\ldots,K\}. τ k = γ T k = τ 1 Ω k − 1 , k ∈ { 1 , … , K } .
Further, the clusters differ in their sizes of the updates per unit activity, and the bounds below depend on these sizes only through their extremes κ − = min l κ l \kappa_-=\min_l\kappa_l κ − = min l κ l and κ + = max l κ l \kappa_+=\max_l\kappa_l κ + = max l κ l .
Proposition 4.10 (Power-law kernel). Let the assumptions of Proposition 4.9 hold with κ l > 0 \kappa_l>0 κ l > 0 for every cluster l l l . Let
I ( t ) = ∫ t / τ K t / τ 1 u α e − u d u , I ∞ = ∫ 0 ∞ u α e − u d u , C ϕ = ∥ ϕ ∥ ln Ω ∑ l = 1 K τ l − α , \mathcal{I}(t)=\int_{t/\tau_K}^{t/\tau_1}u^\alpha e^{-u}\,\mathrm{d}u,\qquad
\mathcal{I}_\infty=\int_0^\infty u^\alpha e^{-u}\,\mathrm{d}u,\qquad
C_\phi=\frac{\lVert\phi\rVert}{\ln\Omega\sum_{l=1}^{K}\tau_l^{-\alpha}}, I ( t ) = ∫ t / τ K t / τ 1 u α e − u d u , I ∞ = ∫ 0 ∞ u α e − u d u , C ϕ = ln Ω ∑ l = 1 K τ l − α ∥ ϕ ∥ , where C ϕ C_\phi C ϕ depends only on ∥ ϕ ∥ \lVert\phi\rVert ∥ ϕ ∥ , α \alpha α , τ 1 \tau_1 τ 1 , Ω \Omega Ω and K K K . Then the following hold.
(i) For every t > 0 t>0 t > 0 ,
κ − κ + I ( t ) Ω 1 + α ≤ t 1 + α ϕ ˉ ( t ) C ϕ ≤ κ + κ − ( Ω 1 + α I ( t ) + ln Ω ( t τ K ) 1 + α e − t / τ K ) . \frac{\kappa_-}{\kappa_+}\frac{\mathcal{I}(t)}{\Omega^{1+\alpha}}
\le\frac{t^{1+\alpha}\bar{\phi}(t)}{C_\phi}
\le\frac{\kappa_+}{\kappa_-}\left(\Omega^{1+\alpha}\mathcal{I}(t)+\ln\Omega\left(\frac{t}{\tau_K}\right)^{1+\alpha}e^{-t/\tau_K}\right). κ + κ − Ω 1 + α I ( t ) ≤ C ϕ t 1 + α ϕ ˉ ( t ) ≤ κ − κ + ( Ω 1 + α I ( t ) + ln Ω ( τ K t ) 1 + α e − t / τ K ) . (ii) Let t t t , τ 1 \tau_1 τ 1 and K K K vary such that τ 1 / t → 0 \tau_1/t\to0 τ 1 / t → 0 and t / τ K → 0 t/\tau_K\to0 t / τ K → 0 , while Ω \Omega Ω , α \alpha α , κ − \kappa_- κ − and κ + \kappa_+ κ + stay fixed. Then
κ − κ + I ∞ Ω 1 + α ≤ lim inf t 1 + α ϕ ˉ ( t ) C ϕ ≤ lim sup t 1 + α ϕ ˉ ( t ) C ϕ ≤ κ + κ − Ω 1 + α I ∞ . \frac{\kappa_-}{\kappa_+}\frac{\mathcal{I}_\infty}{\Omega^{1+\alpha}}
\le\liminf\frac{t^{1+\alpha}\bar{\phi}(t)}{C_\phi}
\le\limsup\frac{t^{1+\alpha}\bar{\phi}(t)}{C_\phi}
\le\frac{\kappa_+}{\kappa_-}\Omega^{1+\alpha}\mathcal{I}_\infty. κ + κ − Ω 1 + α I ∞ ≤ lim inf C ϕ t 1 + α ϕ ˉ ( t ) ≤ lim sup C ϕ t 1 + α ϕ ˉ ( t ) ≤ κ − κ + Ω 1 + α I ∞ .
Between the shortest and the longest horizon, the kernel thus decays as t − 1 − α t^{-1-\alpha} t − 1 − α . As τ 1 − α ≤ ∑ l τ l − α ≤ τ 1 − α / ( 1 − Ω − α ) \tau_1^{-\alpha}\le\sum_l\tau_l^{-\alpha}\le\tau_1^{-\alpha}/(1-\Omega^{-\alpha}) τ 1 − α ≤ ∑ l τ l − α ≤ τ 1 − α / ( 1 − Ω − α ) , the constant C ϕ C_\phi C ϕ stays bounded as K → ∞ K\to\infty K → ∞ . Clusters with longer horizons hence extend the power law to longer times.
Notice that on the logarithmic scale of the horizons the sum in ϕ ˉ \bar{\phi} ϕ ˉ is a Riemann sum,
∑ l = 1 K τ l − 1 − α e − t / τ l ≈ 1 ln Ω ∫ τ 1 τ K τ − 2 − α e − t / τ d τ = t − 1 − α ln Ω I ( t ) . \sum_{l=1}^{K}\tau_l^{-1-\alpha}e^{-t/\tau_l}
\approx\frac{1}{\ln\Omega}\int_{\tau_1}^{\tau_K}\tau^{-2-\alpha}e^{-t/\tau}\,\mathrm{d}\tau
= \frac{t^{-1-\alpha}}{\ln\Omega}\mathcal{I}(t). l = 1 ∑ K τ l − 1 − α e − t / τ l ≈ ln Ω 1 ∫ τ 1 τ K τ − 2 − α e − t / τ d τ = ln Ω t − 1 − α I ( t ) .
Proof. By (4) and Assumption 3.2, the factor γ α \gamma^\alpha γ α cancels in w l = T l − α / ∑ j T j − α w_l=T_l^{-\alpha}/\sum_jT_j^{-\alpha} w l = T l − α / ∑ j T j − α , such that
w l = τ l − α / ∑ j τ j − α . w_l=\tau_l^{-\alpha}/\sum_j\tau_j^{-\alpha}. w l = τ l − α / j ∑ τ j − α .
As κ ˉ = ∑ l κ l w l \bar{\kappa}=\sum_l\kappa_lw_l κ ˉ = ∑ l κ l w l with ∑ l w l = 1 \sum_lw_l=1 ∑ l w l = 1 , we further have κ − ≤ κ ˉ ≤ κ + \kappa_-\le\bar{\kappa}\le\kappa_+ κ − ≤ κ ˉ ≤ κ + . Now, let f ( τ ) = τ − 1 − α e − t / τ f(\tau)=\tau^{-1-\alpha}e^{-t/\tau} f ( τ ) = τ − 1 − α e − t / τ for τ > 0 \tau>0 τ > 0 . Inserting ϖ l = κ l w l / κ ˉ \varpi_l=\kappa_lw_l/\bar{\kappa} ϖ l = κ l w l / κ ˉ into ϕ ˉ \bar{\phi} ϕ ˉ , we obtain
κ − κ + C ϕ ln Ω ∑ l = 1 K f ( τ l ) ≤ ϕ ˉ ( t ) = ∥ ϕ ∥ ∑ l = 1 K κ l κ ˉ τ l − α ∑ j τ j − α e − t / τ l τ l ≤ κ + κ − C ϕ ln Ω ∑ l = 1 K f ( τ l ) . \frac{\kappa_-}{\kappa_+}C_\phi\ln\Omega\sum_{l=1}^{K}f(\tau_l)
\le\bar{\phi}(t)=\lVert\phi\rVert\sum_{l=1}^{K}\frac{\kappa_l}{\bar{\kappa}}\frac{\tau_l^{-\alpha}}{\sum_j\tau_j^{-\alpha}}\frac{e^{-t/\tau_l}}{\tau_l}
\le\frac{\kappa_+}{\kappa_-}C_\phi\ln\Omega\sum_{l=1}^{K}f(\tau_l). κ + κ − C ϕ ln Ω l = 1 ∑ K f ( τ l ) ≤ ϕ ˉ ( t ) = ∥ ϕ ∥ l = 1 ∑ K κ ˉ κ l ∑ j τ j − α τ l − α τ l e − t / τ l ≤ κ − κ + C ϕ ln Ω l = 1 ∑ K f ( τ l ) .
Next, let l ∈ { 1 , … , K − 1 } l \in\{1, \ldots, K-1\} l ∈ { 1 , … , K − 1 } and τ ∈ [ τ l , τ l + 1 ] \tau\in[\tau_l,\tau_{l+1}] τ ∈ [ τ l , τ l + 1 ] . Since τ − 1 − α \tau^{-1-\alpha} τ − 1 − α decreases and e − t / τ e^{-t/\tau} e − t / τ increases in τ \tau τ , we obtain from τ l + 1 = Ω τ l \tau_{l+1}=\Omega\tau_l τ l + 1 = Ω τ l
Ω − 1 − α f ( τ l ) = ( Ω τ l ) − 1 − α e − t / τ l ≤ f ( τ ) ≤ ( τ l + 1 / Ω ) − 1 − α e − t / τ l + 1 = Ω 1 + α f ( τ l + 1 ) . \Omega^{-1-\alpha}f(\tau_l)=(\Omega\tau_l)^{-1-\alpha}e^{-t/\tau_l}\leq f(\tau)\leq(\tau_{l+1}/\Omega)^{-1-\alpha}e^{-t/\tau_{l+1}}=\Omega^{1+\alpha}f(\tau_{l+1}). Ω − 1 − α f ( τ l ) = ( Ω τ l ) − 1 − α e − t / τ l ≤ f ( τ ) ≤ ( τ l + 1 /Ω ) − 1 − α e − t / τ l + 1 = Ω 1 + α f ( τ l + 1 ) .
Integrating against d τ / τ d\tau/\tau d τ / τ over [ τ l , τ l + 1 ] [\tau_l,\tau_{l+1}] [ τ l , τ l + 1 ] , summing over l l l and substituting u = t / τ u=t/\tau u = t / τ , we obtain
ln Ω Ω 1 + α ∑ l = 1 K − 1 f ( τ l ) ≤ ∫ τ 1 τ K f ( τ ) d τ τ = ∫ t / τ K t / τ 1 ( u t ) 1 + α e − u d u u = t − 1 − α I ( t ) ≤ Ω 1 + α ln Ω ∑ l = 2 K f ( τ l ) . \begin{aligned}
\frac{\ln\Omega}{\Omega^{1+\alpha}}\sum_{l=1}^{K-1}f(\tau_l)\leq\int_{\tau_1}^{\tau_K}f(\tau)\frac{d\tau}{\tau}=\int_{t/\tau_K}^{t/\tau_1}\left(\frac{u}{t}\right)^{1+\alpha}e^{-u}\frac{du}{u} \\
&=t^{-1-\alpha}\mathcal{I}(t)\leq\Omega^{1+\alpha}\ln\Omega\sum_{l=2}^{K}f(\tau_l).
\end{aligned} Ω 1 + α ln Ω l = 1 ∑ K − 1 f ( τ l ) ≤ ∫ τ 1 τ K f ( τ ) τ d τ = ∫ t / τ K t / τ 1 ( t u ) 1 + α e − u u d u = t − 1 − α I ( t ) ≤ Ω 1 + α ln Ω l = 2 ∑ K f ( τ l ) .
Adding f ( τ K ) = t − 1 − α ( t / τ K ) 1 + α e − t / τ K f(\tau_K)=t^{-1-\alpha}(t/\tau_K)^{1+\alpha}e^{-t/\tau_K} f ( τ K ) = t − 1 − α ( t / τ K ) 1 + α e − t / τ K to the left sum and f ( τ 1 ) ≥ 0 f(\tau_1)\geq0 f ( τ 1 ) ≥ 0 to the right sum, we obtain
t − 1 − α I ( t ) Ω 1 + α ln Ω ≤ ∑ l = 1 K f ( τ l ) ≤ t − 1 − α ( Ω 1 + α ln Ω I ( t ) + ( t τ K ) 1 + α e − t / τ K ) . \frac{t^{-1-\alpha}\mathcal{I}(t)}{\Omega^{1+\alpha}\ln\Omega}\leq\sum_{l=1}^{K}f(\tau_l)\leq t^{-1-\alpha}\left(\frac{\Omega^{1+\alpha}}{\ln\Omega}\mathcal{I}(t)+\left(\frac{t}{\tau_K}\right)^{1+\alpha}e^{-t/\tau_K}\right). Ω 1 + α ln Ω t − 1 − α I ( t ) ≤ l = 1 ∑ K f ( τ l ) ≤ t − 1 − α ( ln Ω Ω 1 + α I ( t ) + ( τ K t ) 1 + α e − t / τ K ) .
Inserting these bounds into the first display, we obtain (i). For (ii), the limits of integration of I ( t ) \mathcal{I}(t) I ( t ) tend to 0 0 0 and ∞ \infty ∞ , such that I ( t ) → I ∞ \mathcal{I}(t)\to\mathcal{I}_\infty I ( t ) → I ∞ by monotone convergence. Moreover, ( t / τ K ) 1 + α e − t / τ K ≤ ( t / τ K ) 1 + α → 0 (t/\tau_K)^{1+\alpha}e^{-t/\tau_K}\leq(t/\tau_K)^{1+\alpha}\to0 ( t / τ K ) 1 + α e − t / τ K ≤ ( t / τ K ) 1 + α → 0 . Lastly, noting that the bounds in (i) depend on t t t , τ 1 \tau_1 τ 1 and K K K only through I ( t ) \mathcal{I}(t) I ( t ) and t / τ K t/\tau_K t / τ K , we obtain (ii). □
Many desks. We now show that the bounds of Proposition 4.10 (ii) close in a market with many desks, which gives φ ˉ ( t ) / ∥ φ ∥ \bar{\varphi}(t)/\|\varphi\| φ ˉ ( t ) /∥ φ ∥ an exact tail k 0 t − 1 − α k_0t^{-1-\alpha} k 0 t − 1 − α . The two bounds on t 1 + α φ ˉ ( t ) / C φ t^{1+\alpha}\bar{\varphi}(t)/C_\varphi t 1 + α φ ˉ ( t ) / C φ in (ii) differ by the factor
( κ + / κ − ) Ω 1 + α I ∞ ( κ − / κ + ) Ω − 1 − α I ∞ = ( κ + κ − ) 2 Ω 2 ( 1 + α ) > 1. \frac{(\kappa_+/\kappa_-)\Omega^{1+\alpha}\mathcal{I}_\infty}{(\kappa_-/\kappa_+)\Omega^{-1-\alpha}\mathcal{I}_\infty}=\left(\frac{\kappa_+}{\kappa_-}\right)^2\Omega^{2(1+\alpha)}>1. ( κ − / κ + ) Ω − 1 − α I ∞ ( κ + / κ − ) Ω 1 + α I ∞ = ( κ − κ + ) 2 Ω 2 ( 1 + α ) > 1.
Here, the factor Ω 2 ( 1 + α ) \Omega^{2(1+\alpha)} Ω 2 ( 1 + α ) grows with the ratio Ω = τ l + 1 / τ l \Omega=\tau_{l+1}/\tau_l Ω = τ l + 1 / τ l of neighboring horizons, while ( κ + / κ − ) 2 (\kappa_+/\kappa_-)^2 ( κ + / κ − ) 2 grows with the differences between the clusters in the size of the updates per unit activity. Both factors therefore approach one when many clusters with nearly equal horizons and nearly equal sizes of the updates per unit activity fill the range of horizons.
A liquid market comes close to this situation, as many desks with horizons on many scales quote there at the same time. In our model, by contrast, the clusters group these desks into the horizons τ k = τ 1 Ω k − 1 \tau_k=\tau_1\Omega^{k-1} τ k = τ 1 Ω k − 1 , which keep the fixed ratio Ω \Omega Ω between neighbors. We therefore describe a liquid market by the limit of many desks.3 In this limit, the number of clusters grows, K → ∞ K\to\infty K → ∞ , and neighboring horizons approach each other, Ω → 1 + \Omega\to1^+ Ω → 1 + . With N N N desks per cluster and side, the market then holds 2 N K → ∞ 2NK\to\infty 2 N K → ∞ desks. For the second factor ( κ + / κ − ) 2 (\kappa_+/\kappa_-)^2 ( κ + / κ − ) 2 , we note that the protocol treats two clusters alike except for their horizons and volumes, while κ l \kappa_l κ l divides out the volume w l w_l w l . If the targets of all clusters follow the same distribution, clusters with nearly equal horizons thus have nearly equal κ l \kappa_l κ l . In the limit of many desks we therefore let κ l \kappa_l κ l vary continuously with the horizon.
Now, let κ : [ τ 1 , ∞ ) → ( 0 , ∞ ) \kappa:[\tau_1,\infty)\to(0,\infty) κ : [ τ 1 , ∞ ) → ( 0 , ∞ ) be continuous with inf κ > 0 \inf\kappa>0 inf κ > 0 , sup κ < ∞ \sup\kappa<\infty sup κ < ∞ and κ ( τ ) → κ ∞ \kappa(\tau)\to\kappa_\infty κ ( τ ) → κ ∞ as τ → ∞ \tau\to\infty τ → ∞ , and let κ l = κ ( τ l ) \kappa_l=\kappa(\tau_l) κ l = κ ( τ l ) . Further, let
K ( t ) = φ ˉ ( t ) ∥ φ ∥ = ∑ l = 1 K w l e − t / τ l τ l , t ≥ 0 , \mathcal{K}(t)=\frac{\bar{\varphi}(t)}{\|\varphi\|}=\sum_{l=1}^{K}w_l\frac{e^{-t/\tau_l}}{\tau_l},\qquad t\geq0, K ( t ) = ∥ φ ∥ φ ˉ ( t ) = l = 1 ∑ K w l τ l e − t / τ l , t ≥ 0 ,
denote the normalized kernel, which satisfies ∫ 0 ∞ K ( t ) d t = 1 \int_0^\infty\mathcal{K}(t)\,dt=1 ∫ 0 ∞ K ( t ) d t = 1 by Proposition 4.9 . By (4) , the weight ϖ l \varpi_l ϖ l is proportional to κ ( τ l ) τ l − α \kappa(\tau_l)\tau_l^{-\alpha} κ ( τ l ) τ l − α , such that K ( t ) \mathcal{K}(t) K ( t ) is proportional to the sum on the left below. Since ln τ l + 1 − ln τ l = ln Ω \ln\tau_{l+1}-\ln\tau_l=\ln\Omega ln τ l + 1 − ln τ l = ln Ω , we obtain in the limit of many desks
ln Ω ∑ l = 1 K κ ( τ l ) τ l − 1 − α e − t / τ l ⟶ ∫ τ 1 ∞ κ ( τ ) τ − 2 − α e − t / τ d τ = t − 1 − α ∫ 0 t / τ 1 κ ( t u ) u α e − u d u \ln\Omega\sum_{l=1}^{K}\kappa(\tau_l)\tau_l^{-1-\alpha}e^{-t/\tau_l}\longrightarrow\int_{\tau_1}^{\infty}\kappa(\tau)\tau^{-2-\alpha}e^{-t/\tau}\,d\tau=t^{-1-\alpha}\int_0^{t/\tau_1}\kappa\left(\frac{t}{u}\right)u^\alpha e^{-u}\,du ln Ω l = 1 ∑ K κ ( τ l ) τ l − 1 − α e − t / τ l ⟶ ∫ τ 1 ∞ κ ( τ ) τ − 2 − α e − t / τ d τ = t − 1 − α ∫ 0 t / τ 1 κ ( u t ) u α e − u d u
where we substituted u = t / τ u=t/\tau u = t / τ in the last step. As κ ( t / u ) → κ ∞ \kappa(t/u)\to\kappa_\infty κ ( t / u ) → κ ∞ , the last integral converges to κ ∞ I ∞ \kappa_\infty I_\infty κ ∞ I ∞ as t → ∞ t\to\infty t → ∞ .
Proposition 4.11 (Kernel of many desks). Under the assumptions of Proposition 4.10 , let κ l = κ ( τ l ) \kappa_l=\kappa(\tau_l) κ l = κ ( τ l ) for every cluster l l l , and let
ϖ ( τ ) = α τ 1 α κ ( τ ) τ − 1 − α κ ˉ ∞ , κ ˉ ∞ = ∫ τ 1 ∞ α τ 1 α κ ( τ ) τ − 1 − α d τ , \varpi(\tau)=\frac{\alpha\tau_1^\alpha\kappa(\tau)\tau^{-1-\alpha}}{\bar{\kappa}_\infty},\qquad\bar{\kappa}_\infty=\int_{\tau_1}^{\infty}\alpha\tau_1^\alpha\kappa(\tau)\tau^{-1-\alpha}\,d\tau, ϖ ( τ ) = κ ˉ ∞ α τ 1 α κ ( τ ) τ − 1 − α , κ ˉ ∞ = ∫ τ 1 ∞ α τ 1 α κ ( τ ) τ − 1 − α d τ , K ∞ ( t ) = ∫ τ 1 ∞ ϖ ( τ ) e − t / τ τ d τ . \mathcal{K}_\infty(t)=\int_{\tau_1}^{\infty}\varpi(\tau)\frac{e^{-t/\tau}}{\tau}\,d\tau. K ∞ ( t ) = ∫ τ 1 ∞ ϖ ( τ ) τ e − t / τ d τ . Then the following hold.
(i) As Ω → 1 + \Omega\to1^+ Ω → 1 + and K → ∞ K\to\infty K → ∞ with τ K → ∞ \tau_K\to\infty τ K → ∞ , we have K ( t ) → K ∞ ( t ) \mathcal{K}(t)\to\mathcal{K}_\infty(t) K ( t ) → K ∞ ( t ) for every t > 0 t>0 t > 0 , and
∫ 0 ∞ K ∞ ( t ) d t = 1. \int_0^\infty\mathcal{K}_\infty(t)\,dt=1. ∫ 0 ∞ K ∞ ( t ) d t = 1. (ii) As t → ∞ t\to\infty t → ∞ ,
t 1 + α K ∞ ( t ) → k 0 , α x α ∫ x ∞ K ∞ ( t ) d t → k 0 as x → ∞ , k 0 = α τ 1 α κ ∞ I ∞ κ ˉ ∞ . t^{1+\alpha}\mathcal{K}_\infty(t)\to k_0,\qquad\alpha x^\alpha\int_x^\infty\mathcal{K}_\infty(t)\,dt\to k_0\quad\text{as }x\to\infty,\qquad k_0=\frac{\alpha\tau_1^\alpha\kappa_\infty I_\infty}{\bar{\kappa}_\infty}. t 1 + α K ∞ ( t ) → k 0 , α x α ∫ x ∞ K ∞ ( t ) d t → k 0 as x → ∞ , k 0 = κ ˉ ∞ α τ 1 α κ ∞ I ∞ .
The exponent of the tail is thus set by the volumes (4) alone, while the sizes of the updates per unit activity enter only through the constant k 0 k_0 k 0 . If all clusters have the same size of the updates per unit activity, κ = κ ∞ \kappa=\kappa_\infty κ = κ ∞ , then κ ˉ ∞ = κ ∞ \bar{\kappa}_\infty=\kappa_\infty κ ˉ ∞ = κ ∞ , such that k 0 = α τ 1 α I ∞ k_0=\alpha\tau_1^\alpha I_\infty k 0 = α τ 1 α I ∞ depends only on α \alpha α and the shortest horizon.
Proof. For (i), let g g g be continuous on [ τ 1 , ∞ ) [\tau_1,\infty) [ τ 1 , ∞ ) with ∣ g ( τ ) ∣ ≤ sup κ τ − α |g(\tau)|\le\sup\kappa\tau^{-\alpha} ∣ g ( τ ) ∣ ≤ sup κ τ − α . We first show that
ln Ω ∑ l = 1 K g ( τ l ) ⟶ ∫ τ 1 ∞ g ( τ ) d τ τ as Ω → 1 + , τ K → ∞ . \ln\Omega\sum_{l=1}^{K}g(\tau_l)\longrightarrow\int_{\tau_1}^{\infty}g(\tau)\frac{d\tau}{\tau}\quad\text{as }\Omega\to1^+,\ \tau_K\to\infty. ln Ω l = 1 ∑ K g ( τ l ) ⟶ ∫ τ 1 ∞ g ( τ ) τ d τ as Ω → 1 + , τ K → ∞.
As ln τ l + 1 − ln τ l = ln Ω \ln\tau_{l+1}-\ln\tau_l=\ln\Omega ln τ l + 1 − ln τ l = ln Ω , the left side is a Riemann sum of the function ln τ ↦ g ( τ ) \ln\tau\mapsto g(\tau) ln τ ↦ g ( τ ) with the step ln Ω \ln\Omega ln Ω . For a horizon τ ^ > τ 1 \hat{\tau}>\tau_1 τ ^ > τ 1 , the terms with τ l ≤ τ ^ \tau_l\le\hat{\tau} τ l ≤ τ ^ thus converge to ∫ τ 1 τ ^ g ( τ ) d τ / τ \int_{\tau_1}^{\hat{\tau}}g(\tau)\,d\tau/\tau ∫ τ 1 τ ^ g ( τ ) d τ / τ by uniform continuity on [ ln τ 1 , ln τ ^ ] [\ln\tau_1,\ln\hat{\tau}] [ ln τ 1 , ln τ ^ ] . The remaining terms and the remaining integral satisfy
ln Ω ∑ l : τ l > τ ^ ∣ g ( τ l ) ∣ ≤ sup κ ln Ω ∑ j ∈ N ( τ ^ Ω j ) − α = sup κ ln Ω 1 − Ω − α τ ^ − α , \ln\Omega\sum_{l:\tau_l>\hat{\tau}}|g(\tau_l)|\le\sup\kappa\ln\Omega\sum_{j\in\mathbb{N}}(\hat{\tau}\Omega^j)^{-\alpha}=\sup\kappa\frac{\ln\Omega}{1-\Omega^{-\alpha}}\hat{\tau}^{-\alpha}, ln Ω l : τ l > τ ^ ∑ ∣ g ( τ l ) ∣ ≤ sup κ ln Ω j ∈ N ∑ ( τ ^ Ω j ) − α = sup κ 1 − Ω − α ln Ω τ ^ − α ,
∫ τ ^ ∞ ∣ g ( τ ) ∣ d τ τ ≤ sup κ τ ^ − α α , \int_{\hat{\tau}}^\infty|g(\tau)|\frac{d\tau}{\tau}\le\sup\kappa\frac{\hat{\tau}^{-\alpha}}{\alpha}, ∫ τ ^ ∞ ∣ g ( τ ) ∣ τ d τ ≤ sup κ α τ ^ − α ,
where ln Ω / ( 1 − Ω − α ) → 1 / α \ln\Omega/(1-\Omega^{-\alpha})\to1/\alpha ln Ω/ ( 1 − Ω − α ) → 1/ α . Letting first Ω → 1 + \Omega\to1^+ Ω → 1 + and then τ ^ → ∞ \hat{\tau}\to\infty τ ^ → ∞ , we obtain the claim. Next, by (4) and Assumption 3.2 we have ϖ l = κ l τ l − α / ∑ j κ j τ j − α \varpi_l=\kappa_l\tau_l^{-\alpha}/\sum_j\kappa_j\tau_j^{-\alpha} ϖ l = κ l τ l − α / ∑ j κ j τ j − α , such that
K ( t ) = ln Ω ∑ l κ ( τ l ) τ l − α e − t / τ l / τ l ln Ω ∑ j κ ( τ j ) τ j − α ⟶ ∫ τ 1 ∞ κ ( τ ) τ − 2 − α e − t / τ d τ ∫ τ 1 ∞ κ ( τ ) τ − 1 − α d τ = K ∞ ( t ) , \mathcal{K}(t)=\frac{\ln\Omega\sum_l\kappa(\tau_l)\tau_l^{-\alpha}e^{-t/\tau_l}/\tau_l}{\ln\Omega\sum_j\kappa(\tau_j)\tau_j^{-\alpha}}\longrightarrow\frac{\int_{\tau_1}^{\infty}\kappa(\tau)\tau^{-2-\alpha}e^{-t/\tau}\,d\tau}{\int_{\tau_1}^{\infty}\kappa(\tau)\tau^{-1-\alpha}\,d\tau}=\mathcal{K}_\infty(t), K ( t ) = ln Ω ∑ j κ ( τ j ) τ j − α ln Ω ∑ l κ ( τ l ) τ l − α e − t / τ l / τ l ⟶ ∫ τ 1 ∞ κ ( τ ) τ − 1 − α d τ ∫ τ 1 ∞ κ ( τ ) τ − 2 − α e − t / τ d τ = K ∞ ( t ) ,
where we applied the claim to g ( τ ) = κ ( τ ) τ − 1 − α e − t / τ g(\tau)=\kappa(\tau)\tau^{-1-\alpha}e^{-t/\tau} g ( τ ) = κ ( τ ) τ − 1 − α e − t / τ and to g ( τ ) = κ ( τ ) τ − α g(\tau)=\kappa(\tau)\tau^{-\alpha} g ( τ ) = κ ( τ ) τ − α . Moreover, by Fubini’s theorem and ∫ 0 ∞ e − t / τ d t / τ = 1 \int_{0}^{\infty}e^{-t/\tau}\,dt/\tau=1 ∫ 0 ∞ e − t / τ d t / τ = 1 , we obtain ∫ 0 ∞ K ∞ ( t ) d t = ∫ τ 1 ∞ ϖ ( τ ) d τ = 1 \int_{0}^{\infty}K_{\infty}(t)\,dt=\int_{\tau_{1}}^{\infty}\varpi(\tau)\,d\tau=1 ∫ 0 ∞ K ∞ ( t ) d t = ∫ τ 1 ∞ ϖ ( τ ) d τ = 1 . For (ii), substituting u = t / τ u=t/\tau u = t / τ , we obtain
t 1 + α K ∞ ( t ) = α τ 1 α κ ˉ ∞ ∫ 0 t / τ 1 κ ( t u ) u α e − u d u . t^{1+\alpha}K_{\infty}(t)=\frac{\alpha\tau_{1}^{\alpha}}{\bar{\kappa}_{\infty}}\int_{0}^{t/\tau_{1}}\kappa\left(\frac{t}{u}\right)u^{\alpha}e^{-u}\,du. t 1 + α K ∞ ( t ) = κ ˉ ∞ α τ 1 α ∫ 0 t / τ 1 κ ( u t ) u α e − u d u .
For every u > 0 u>0 u > 0 , κ ( t / u ) → κ ∞ \kappa(t/u)\to\kappa_{\infty} κ ( t / u ) → κ ∞ as t → ∞ t\to\infty t → ∞ , while the integrand is bounded by sup κ u α e − u \sup\kappa\,u^{\alpha}e^{-u} sup κ u α e − u . By dominated convergence, t 1 + α K ∞ ( t ) → k 0 t^{1+\alpha}K_{\infty}(t)\to k_{0} t 1 + α K ∞ ( t ) → k 0 , that is, K ∞ ( t ) = t − 1 − α ( k 0 + o ( 1 ) ) K_{\infty}(t)=t^{-1-\alpha}(k_{0}+o(1)) K ∞ ( t ) = t − 1 − α ( k 0 + o ( 1 )) . Lastly, integrating from x x x to ∞ \infty ∞ , we obtain
α x α ∫ x ∞ K ∞ ( t ) d t = α x α ∫ x ∞ t − 1 − α ( k 0 + o ( 1 ) ) d t = k 0 + o ( 1 ) as x → ∞ . \alpha x^{\alpha}\int_{x}^{\infty}K_{\infty}(t)\,dt=\alpha x^{\alpha}\int_{x}^{\infty}t^{-1-\alpha}(k_{0}+o(1))\,dt=k_{0}+o(1)\qquad\text{as }x\to\infty. α x α ∫ x ∞ K ∞ ( t ) d t = α x α ∫ x ∞ t − 1 − α ( k 0 + o ( 1 )) d t = k 0 + o ( 1 ) as x → ∞.
Spacing of the horizons. Proposition 4.11 closes the bounds of Proposition 4.10 only in the limit Ω → 1 + \Omega\to1^{+} Ω → 1 + . We now show that already a few clusters per decade of horizons make the kernel a power law to high accuracy. To isolate the role of the ratio Ω \Omega Ω , we let all clusters have the same size of the updates per unit activity and let the horizons τ k = τ 1 Ω k − 1 \tau_k=\tau_1\Omega^{k-1} τ k = τ 1 Ω k − 1 run over all k ∈ Z k\in\mathbb{Z} k ∈ Z . A market with K K K clusters drops the terms with k ≤ 0 k\leq0 k ≤ 0 and k > K k>K k > K . As in the proof of Proposition 4.10 (ii), these terms vanish as τ 1 / t → 0 \tau_1/t\to0 τ 1 / t → 0 and t / τ K → 0 t/\tau_K\to0 t / τ K → 0 . Now, let
J m = ∫ 0 ∞ u α + 2 π i m / ln Ω e − u d u , m ∈ N , \mathcal{J}_m=\int_{0}^{\infty}u^{\alpha+2\pi im/\ln\Omega}e^{-u}\,du,\qquad m\in\mathbb{N}, J m = ∫ 0 ∞ u α + 2 π im / l n Ω e − u d u , m ∈ N ,
which extend I ∞ \mathcal{I}_{\infty} I ∞ to complex powers of u u u , such that J 0 = I ∞ \mathcal{J}_0=\mathcal{I}_{\infty} J 0 = I ∞ .
Proof. Let F ( z ) = exp ( ( 1 + α ) z − e z ) F(z)=\exp((1+\alpha)z-e^{z}) F ( z ) = exp (( 1 + α ) z − e z ) for z ∈ R z\in\mathbb{R} z ∈ R . Writing y = ln ( t / τ 1 ) y=\ln(t/\tau_1) y = ln ( t / τ 1 ) and ln τ k = ln τ 1 + ( k − 1 ) ln Ω \ln\tau_k=\ln\tau_1+(k-1)\ln\Omega ln τ k = ln τ 1 + ( k − 1 ) ln Ω , we obtain
t 1 + α ∑ k ∈ Z τ k − 1 − α e − t / τ k = ∑ k ∈ Z F ( y − ( k − 1 ) ln Ω ) = ∑ k ∈ Z F ( y − k ln Ω ) . t^{1+\alpha}\sum_{k\in\mathbb{Z}}\tau_k^{-1-\alpha}e^{-t/\tau_k}=\sum_{k\in\mathbb{Z}}F(y-(k-1)\ln\Omega)=\sum_{k\in\mathbb{Z}}F(y-k\ln\Omega). t 1 + α k ∈ Z ∑ τ k − 1 − α e − t / τ k = k ∈ Z ∑ F ( y − ( k − 1 ) ln Ω ) = k ∈ Z ∑ F ( y − k ln Ω ) .
Substituting u = e z u=e^{z} u = e z , we obtain for the Fourier transform of F F F
F ^ ( ξ ) = ∫ R F ( z ) e − i ξ z d z = ∫ 0 ∞ u α − i ξ e − u d u . \widehat{F}(\xi)=\int_{\mathbb{R}}F(z)e^{-i\xi z}\,dz=\int_{0}^{\infty}u^{\alpha-i\xi}e^{-u}\,du. F ( ξ ) = ∫ R F ( z ) e − i ξ z d z = ∫ 0 ∞ u α − i ξ e − u d u .
such that F ^ ( 0 ) = I ∞ \widehat{F}(0)=\mathcal{I}_{\infty} F ( 0 ) = I ∞ , F ^ ( 2 π m / ln Ω ) = J m \widehat{F}(2\pi m/\ln\Omega)=\mathcal{J}_{m} F ( 2 π m / ln Ω ) = J m and F ^ ( − ξ ) = F ^ ( ξ ) \widehat{F}(-\xi)=\widehat{F}(\xi) F ( − ξ ) = F ( ξ ) . As F F F decays exponentially, and F ^ \widehat{F} F by Stirling’s formula, we obtain from the Poisson summation formula
∑ k ∈ Z F ( y − k ln Ω ) = 1 ln Ω ∑ m ∈ Z F ^ ( 2 π m ln Ω ) e 2 π i m y / ln Ω = 1 ln Ω ( I ∞ + 2 ∑ m ≥ 1 Re ( J m e 2 π i m y / ln Ω ) ) . \sum_{k\in\mathbb{Z}}F(y-k\ln\Omega)=\frac{1}{\ln\Omega}\sum_{m\in\mathbb{Z}}\widehat{F}\left(\frac{2\pi m}{\ln\Omega}\right)e^{2\pi imy/\ln\Omega}=\frac{1}{\ln\Omega}\left(\mathcal{I}_{\infty}+2\sum_{m\geq1}\operatorname{Re}\left(\mathcal{J}_{m}e^{2\pi imy/\ln\Omega}\right)\right). k ∈ Z ∑ F ( y − k ln Ω ) = ln Ω 1 m ∈ Z ∑ F ( ln Ω 2 π m ) e 2 π im y / l n Ω = ln Ω 1 ( I ∞ + 2 m ≥ 1 ∑ Re ( J m e 2 π im y / l n Ω ) ) .
where we combined the terms m m m and − m -m − m . Multiplying by ln Ω / I ∞ \ln\Omega/\mathcal{I}_{\infty} ln Ω/ I ∞ , we obtain the identity. Moreover, using ∣ Re ( z ) ∣ ≤ ∣ z ∣ |\operatorname{Re}(z)|\leq|z| ∣ Re ( z ) ∣ ≤ ∣ z ∣ , we obtain the bound. Lastly, with y m = 2 π m / ln Ω y_m=2\pi m/\ln\Omega y m = 2 π m / ln Ω , the product formula for ∣ J m ∣ |\mathcal{J}_m| ∣ J m ∣ gives
∣ J m ∣ 2 I ∞ 2 = ∏ n ∈ N ( 1 + y m 2 ( 1 + α + n ) 2 ) − 1 . \frac{|\mathcal{J}_{m}|^{2}}{\mathcal{I}_{\infty}^{2}}=\prod_{n\in\mathbb{N}}\left(1+\frac{y_m^{2}}{(1+\alpha+n)^{2}}\right)^{-1}. I ∞ 2 ∣ J m ∣ 2 = n ∈ N ∏ ( 1 + ( 1 + α + n ) 2 y m 2 ) − 1 .
The right side decreases in y m y_m y m and thus increases in Ω \Omega Ω . □
We next combine the power law t − 1 − α t^{-1-\alpha} t − 1 − α with ∥ ϕ ∥ \|\phi\| ∥ ϕ ∥ close to one to obtain the volatility of the mid price.
Roughness
By Proposition 4.11 , a market with many desks has a normalized kernel with the tail k 0 t − 1 − α k_0t^{-1-\alpha} k 0 t − 1 − α . By Proposition 4.8 , its norm ∥ ϕ ∥ \|\phi\| ∥ ϕ ∥ can approach one only if p 0 p_0 p 0 vanishes, as p 0 ≤ p ‾ ( 1 − ∥ ϕ ∥ ) p_0\leq\overline{p}(1-\|\phi\|) p 0 ≤ p ( 1 − ∥ ϕ ∥ ) . In this section we let ∥ ϕ ∥ \|\phi\| ∥ ϕ ∥ approach one and show that the expected size of the updates is then rough with the Hurst parameter H = α − 1 / 2 H=\alpha-1/2 H = α − 1/2 . As the size of the updates moves the mid price, the volatility of the mid price is then rough as well.
Self-excitation close to the threshold
By Proposition 4.11 , the normalized kernel of a market with many desks has the tail
lim x → ∞ α x α ∫ x ∞ K ∞ ( t ) d t = k 0 . \lim_{x\to\infty}\alpha x^{\alpha}\int_{x}^{\infty}\mathcal{K}_{\infty}(t)\,dt=k_0. x → ∞ lim α x α ∫ x ∞ K ∞ ( t ) d t = k 0 .
For α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , the Karamata Tauberian theorem [6 Corollary 8.1.7] turns this tail into the behavior of the Laplace transform of K ∞ \mathcal{K}_{\infty} K ∞ near zero,
K ^ ∞ ( z ) = ∫ 0 ∞ e − z t K ∞ ( t ) d t = 1 − C 0 z α + o ( z α ) as z → 0 + , C 0 = k 0 α ∫ 0 ∞ u − α e − u d u . \widehat{\mathcal{K}}_{\infty}(z)=\int_{0}^{\infty}e^{-zt}\mathcal{K}_{\infty}(t)\,dt=1-C_0z^{\alpha}+o(z^{\alpha})\text{ as }z\to0^{+},\qquad C_0=\frac{k_0}{\alpha}\int_{0}^{\infty}u^{-\alpha}e^{-u}\,du. K ∞ ( z ) = ∫ 0 ∞ e − z t K ∞ ( t ) d t = 1 − C 0 z α + o ( z α ) as z → 0 + , C 0 = α k 0 ∫ 0 ∞ u − α e − u d u .
As in Corollary 4.6 , the total response of the kernel ∥ ϕ ∥ K ∞ \|\phi\|\mathcal{K}_{\infty} ∥ ϕ ∥ K ∞ thus has the Laplace transform
R ^ ( z ) = ∥ ϕ ∥ K ^ ∞ ( z ) 1 − ∥ ϕ ∥ K ^ ∞ ( z ) = ∥ ϕ ∥ ( 1 + o ( 1 ) ) ( 1 − ∥ ϕ ∥ ) ( 1 + ( T z ) α ) + o ( z α ) , T = ( ∥ ϕ ∥ C 0 1 − ∥ ϕ ∥ ) 1 / α . \widehat{R}(z)=\frac{\|\phi\|\widehat{\mathcal{K}}_{\infty}(z)}{1-\|\phi\|\widehat{\mathcal{K}}_{\infty}(z)}=\frac{\|\phi\|(1+o(1))}{(1-\|\phi\|)(1+(\mathcal{T}z)^{\alpha})+o(z^{\alpha})},\qquad\mathcal{T}=\left(\frac{\|\phi\|C_0}{1-\|\phi\|}\right)^{1/\alpha}. R ( z ) = 1 − ∥ ϕ ∥ K ∞ ( z ) ∥ ϕ ∥ K ∞ ( z ) = ( 1 − ∥ ϕ ∥ ) ( 1 + ( T z ) α ) + o ( z α ) ∥ ϕ ∥ ( 1 + o ( 1 )) , T = ( 1 − ∥ ϕ ∥ ∥ ϕ ∥ C 0 ) 1/ α .
As R ^ ( z ) \widehat{R}(z) R ( z ) weights the response up to times of order 1 / z 1/z 1/ z , the response reaches its total ∥ ϕ ∥ / ( 1 − ∥ ϕ ∥ ) \|\phi\|/(1-\|\phi\|) ∥ ϕ ∥/ ( 1 − ∥ ϕ ∥ ) only over times longer than T \mathcal{T} T . This time diverges as ∥ ϕ ∥ → 1 − \|\phi\|\to1^{-} ∥ ϕ ∥ → 1 − . By Proposition 4.8 , the mean activity tends to p 0 / ( 1 − ∥ ϕ ∥ ) p_0/(1-\|\phi\|) p 0 / ( 1 − ∥ ϕ ∥ ) . We therefore measure time in units of T \mathcal{T} T and the activity in units of its mean.
On times shorter than T \mathcal{T} T , the response has not yet reached its total. For 1 / T ≪ z ≪ 1 / τ 1 1/\mathcal{T}\ll z\ll1/\tau_1 1/ T ≪ z ≪ 1/ τ 1 , the Laplace transform R ^ ( z ) \widehat{R}(z) R ( z ) becomes
R ^ ( z ) ≈ ∥ ϕ ∥ ( 1 − ∥ ϕ ∥ ) ( T z ) α = 1 C 0 z α = ∫ 0 ∞ e − z t t α − 1 C 0 ∫ 0 ∞ u α − 1 e − u d u d t , \widehat{R}(z)\approx\frac{\|\phi\|}{(1-\|\phi\|)(\mathcal{T}z)^{\alpha}}=\frac{1}{C_0z^{\alpha}}=\int_{0}^{\infty}e^{-zt}\frac{t^{\alpha-1}}{C_0\int_{0}^{\infty}u^{\alpha-1}e^{-u}\,du}\,dt, R ( z ) ≈ ( 1 − ∥ ϕ ∥ ) ( T z ) α ∥ ϕ ∥ = C 0 z α 1 = ∫ 0 ∞ e − z t C 0 ∫ 0 ∞ u α − 1 e − u d u t α − 1 d t ,
such that the response decays as t α − 1 t^{\alpha-1} t α − 1 on these times. Now, let I ~ n = I n − E [ I n ∣ F n ] \widetilde{I}_n=I_n-\mathbb{E}[I_n\mid\mathcal{F}_n] I n = I n − E [ I n ∣ F n ] denote the deviation of the size of the updates from its expectation, which has the mean zero given F n \mathcal{F}_n F n . Solving the self-excitation (Proposition 4.5 ) as in the proof of Proposition 4.8 , we obtain
E [ I n ∣ F n ] = p 0 κ ˉ ( 1 + ∑ s = 1 n R ( s ) ) + ∑ m = 0 n − 1 R ( n − m ) I ~ m . \mathbb{E}[I_n\mid\mathcal{F}_n]=p_0\bar{\kappa}\left(1+\sum_{s=1}^{n}R(s)\right)+\sum_{m=0}^{n-1}R(n-m)\widetilde{I}_m. E [ I n ∣ F n ] = p 0 κ ˉ ( 1 + s = 1 ∑ n R ( s ) ) + m = 0 ∑ n − 1 R ( n − m ) I m .
On times shorter than T \mathcal{T} T , the last sum thus weights each deviation I ~ m \widetilde{I}_m I m with ( n − m ) α − 1 (n-m)^{\alpha-1} ( n − m ) α − 1 up to a constant factor. We next state the limit of this representation for self-excitation in continuous time.
In a market that runs in continuous time, the desks act at random times. The number N t T N_t^{\mathcal{T}} N t T of their actions up to the time t ∈ [ 0 , T ] t\in[0,\mathcal{T}] t ∈ [ 0 , T ] then starts at zero and increases by one at each action. Let F t T \mathcal{F}_t^{\mathcal{T}} F t T be the σ \sigma σ -algebra generated by ( N s T ) s ≤ t (N_s^{\mathcal{T}})_{s\leq t} ( N s T ) s ≤ t . The conditional expectation E [ Y n ∣ F n ] \mathbb{E}[Y_n\mid\mathcal{F}_n] E [ Y n ∣ F n ] (Definition 4.1 ) then becomes the intensity λ t T ≥ 0 \lambda_t^{\mathcal{T}}\geq0 λ t T ≥ 0 , the adapted process with
E [ ( N b T − N a T ) 1 A ] = E [ ∫ a b λ s T 1 A d s ] , 0 ≤ a < b ≤ T , A ∈ F a T . \mathbb{E}\left[(N_b^{\mathcal{T}}-N_a^{\mathcal{T}})\mathbf{1}_A\right]=\mathbb{E}\left[\int_a^b\lambda_s^{\mathcal{T}}\mathbf{1}_A\,\mathrm{d}s\right],\qquad0\leq a<b\leq\mathcal{T},\quad A\in\mathcal{F}_a^{\mathcal{T}}. E [ ( N b T − N a T ) 1 A ] = E [ ∫ a b λ s T 1 A d s ] , 0 ≤ a < b ≤ T , A ∈ F a T .
Definition 5.1 (Hawkes process). The number of actions N T N^{\mathcal{T}} N T is a Hawkes process with the base rate μ T > 0 \mu_{\mathcal{T}}>0 μ T > 0 and the kernel K T : [ 0 , ∞ ) → [ 0 , ∞ ) \mathcal{K}_{\mathcal{T}}:[0,\infty)\to[0,\infty) K T : [ 0 , ∞ ) → [ 0 , ∞ ) with ∥ K T ∥ = ∫ 0 ∞ K T ( s ) d s < 1 \lVert\mathcal{K}_{\mathcal{T}}\rVert=\int_0^\infty\mathcal{K}_{\mathcal{T}}(s)\,\mathrm{d}s<1 ∥ K T ∥ = ∫ 0 ∞ K T ( s ) d s < 1 if its intensity is
λ t T = μ T + ∫ 0 t K T ( t − s ) d N s T , t ∈ [ 0 , T ] . \lambda_t^{\mathcal{T}}=\mu_{\mathcal{T}}+\int_0^t\mathcal{K}_{\mathcal{T}}(t-s)\,\mathrm{d}N_s^{\mathcal{T}},\qquad t\in[0,\mathcal{T}]. λ t T = μ T + ∫ 0 t K T ( t − s ) d N s T , t ∈ [ 0 , T ] .
Let R T = ∑ j ∈ N ∖ { 0 } K ∗ j R_{\mathcal{T}}=\sum_{j\in\mathbb{N}\setminus\{0\}}\mathcal{K}^{*j} R T = ∑ j ∈ N ∖ { 0 } K ∗ j denote the total response of the Hawkes process, and let M t T = N t T − ∫ 0 t λ s T d s M_t^{\mathcal{T}}=N_t^{\mathcal{T}}-\int_0^t\lambda_s^{\mathcal{T}}\,\mathrm{d}s M t T = N t T − ∫ 0 t λ s T d s denote the deviation of the number of actions from its integrated intensity, which is a martingale. Solving the self-excitation as for the expected size of the updates, we obtain
λ t T = μ T ( 1 + ∫ 0 t R T ( s ) d s ) + ∫ 0 t R T ( t − s ) d M s T . \lambda_t^{\mathcal{T}}=\mu_{\mathcal{T}}\left(1+\int_0^tR_{\mathcal{T}}(s)\,\mathrm{d}s\right)+\int_0^tR_{\mathcal{T}}(t-s)\,\mathrm{d}M_s^{\mathcal{T}}. λ t T = μ T ( 1 + ∫ 0 t R T ( s ) d s ) + ∫ 0 t R T ( t − s ) d M s T .
The time scale of the response satisfies T α ( 1 − ∥ φ ∥ ) = ∥ φ ∥ C 0 \mathcal{T}^{\alpha}(1-\lVert\varphi\rVert)=\lVert\varphi\rVert C_0 T α ( 1 − ∥ φ ∥) = ∥ φ ∥ C 0 . As T → ∞ \mathcal{T}\to\infty T → ∞ , the norm of the kernel thus approaches one at the rate T − α \mathcal{T}^{-\alpha} T − α . We therefore consider, for every T > 0 \mathcal{T}>0 T > 0 , a Hawkes process N T N^{\mathcal{T}} N T with the kernel K T = ∥ K T ∥ K \mathcal{K}_{\mathcal{T}}=\lVert\mathcal{K}_{\mathcal{T}}\rVert\mathcal{K} K T = ∥ K T ∥ K and ∥ K T ∥ → 1 − \lVert\mathcal{K}_{\mathcal{T}}\rVert\to1^- ∥ K T ∥ → 1 − . As K ∞ \mathcal{K}_\infty K ∞ does by Proposition 4.11 , the kernel K \mathcal{K} K with ∥ K ∥ = 1 \lVert\mathcal{K}\rVert=1 ∥ K ∥ = 1 has a tail
lim x → ∞ α x α ∫ x ∞ K ( s ) d s = k 0 . \lim_{x\to\infty}\alpha x^\alpha\int_x^\infty\mathcal{K}(s)\,\mathrm{d}s=k_0. x → ∞ lim α x α ∫ x ∞ K ( s ) d s = k 0 .
In units of T \mathcal{T} T and of the mean intensity μ T / ( 1 − ∥ K T ∥ ) \mu_{\mathcal{T}}/(1-\lVert\mathcal{K}_{\mathcal{T}}\rVert) μ T / ( 1 − ∥ K T ∥) , the number of actions becomes
λ t T = 1 − ∥ K T ∥ T μ T N t T , t ∈ [ 0 , 1 ] . \lambda_t^{\mathcal{T}}=\frac{1-\lVert\mathcal{K}_{\mathcal{T}}\rVert}{\mathcal{T}\mu_{\mathcal{T}}}N_t^{\mathcal{T}},\qquad t\in[0,1]. λ t T = T μ T 1 − ∥ K T ∥ N t T , t ∈ [ 0 , 1 ] .
On the same time scale, the total response R T R_{\mathcal{T}} R T normalized to the total weight one is ρ T ( x ) = T R T ( T x ) / ∥ R T ∥ \rho^{\mathcal{T}}(x)=\mathcal{T}R_{\mathcal{T}}(\mathcal{T}x)/\lVert R_{\mathcal{T}}\rVert ρ T ( x ) = T R T ( T x ) / ∥ R T ∥ . By the Karamata Tauberian theorem, this tail gives the Laplace transform of K \mathcal{K} K near zero with the same constant C 0 C_0 C 0 as for K ∞ \mathcal{K}_\infty K ∞ ,
K ^ ( z ) = ∫ 0 ∞ e − z t K ( t ) d t = 1 − C 0 z α + o ( z α ) as z → 0 + . \widehat{\mathcal{K}}(z)=\int_0^\infty e^{-zt}\mathcal{K}(t)\,\mathrm{d}t=1-C_0z^\alpha+o(z^\alpha)\qquad\text{as }z\to0^+. K ( z ) = ∫ 0 ∞ e − z t K ( t ) d t = 1 − C 0 z α + o ( z α ) as z → 0 + .
Inserting this expansion, we find
∫ 0 ∞ e − z x ρ T ( x ) d x = ( 1 − ∥ K T ∥ ) K ^ ( z / T ) 1 − ∥ K T ∥ K ^ ( z / T ) → C 1 C 1 + z α if T α ( 1 − ∥ K T ∥ ) C 0 → C 1 > 0. \int_0^\infty e^{-zx}\rho^{\mathcal T}(x)\,dx=\frac{(1-\lVert K_{\mathcal T}\rVert)\widehat K(z/\mathcal T)}{1-\lVert K_{\mathcal T}\rVert\widehat K(z/\mathcal T)}\to\frac{C_1}{C_1+z^\alpha}\quad\text{if}\quad\frac{\mathcal T^\alpha(1-\lVert K_{\mathcal T}\rVert)}{C_0}\to C_1>0. ∫ 0 ∞ e − z x ρ T ( x ) d x = 1 − ∥ K T ∥ K ( z / T ) ( 1 − ∥ K T ∥) K ( z / T ) → C 1 + z α C 1 if C 0 T α ( 1 − ∥ K T ∥) → C 1 > 0.
We let G \mathcal G G denote the function on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) with this limit as its Laplace transform.
Theorem 5.2 (Hawkes processes close to the threshold [41 Theorems 3.1 and 3.2] ). For every T > 0 \mathcal T>0 T > 0 , let N T N^{\mathcal T} N T be a Hawkes process with the base rate μ T \mu_{\mathcal T} μ T and the kernel ∥ K T ∥ K \lVert K_{\mathcal T}\rVert K ∥ K T ∥ K , where ∥ K ∥ = 1 \lVert K\rVert=1 ∥ K ∥ = 1 . Assume that
(J1) lim x → ∞ α x α ∫ x ∞ K ( s ) d s = k 0 \lim_{x\to\infty}\alpha x^\alpha\int_x^\infty K(s)\,ds=k_0 lim x → ∞ α x α ∫ x ∞ K ( s ) d s = k 0 for some α ∈ ( 1 / 2 , 1 ) \alpha\in(1/2,1) α ∈ ( 1/2 , 1 ) and k 0 > 0 k_0>0 k 0 > 0 ;
(J2) T α ( 1 − ∥ K T ∥ ) → C 1 C 0 \mathcal T^\alpha(1-\lVert K_{\mathcal T}\rVert)\to C_1C_0 T α ( 1 − ∥ K T ∥) → C 1 C 0 and T 1 − α μ T → C 2 / C 0 \mathcal T^{1-\alpha}\mu_{\mathcal T}\to C_2/C_0 T 1 − α μ T → C 2 / C 0 as T → ∞ \mathcal T\to\infty T → ∞ , for some C 1 , C 2 > 0 C_1,C_2>0 C 1 , C 2 > 0 .
Then every sequence T k → ∞ \mathcal T_k\to\infty T k → ∞ has a subsequence along which X T k X^{\mathcal T_k} X T k converges in law. Every such limit X X X is differentiable, and there is a Brownian motion B B B such that
X t ′ = ∫ 0 t G ( s ) d s + 1 C 1 C 2 ∫ 0 t G ( t − s ) X s ′ d B s , t ∈ [ 0 , 1 ] . X'_t=\int_0^t\mathcal G(s)\,ds+\frac{1}{\sqrt{C_1C_2}}\int_0^t\mathcal G(t-s)\sqrt{X'_s}\,dB_s,\qquad t\in[0,1]. X t ′ = ∫ 0 t G ( s ) d s + C 1 C 2 1 ∫ 0 t G ( t − s ) X s ′ d B s , t ∈ [ 0 , 1 ] . Moreover, for every h < α − 1 / 2 h<\alpha-1/2 h < α − 1/2 ,
sup 0 ≤ s < t ≤ 1 ∣ X t ′ − X s ′ ∣ ( t − s ) h < ∞ almost surely . \sup_{0\leq s<t\leq1}\frac{\lvert X'_t-X'_s\rvert}{(t-s)^h}<\infty\qquad\text{almost surely}. 0 ≤ s < t ≤ 1 sup ( t − s ) h ∣ X t ′ − X s ′ ∣ < ∞ almost surely .
Our normalization of X T X^{\mathcal T} X T differs from that in [41 ] by the factor C 0 T 1 − α μ T / C 2 C_0\mathcal T^{1-\alpha}\mu_{\mathcal T}/C_2 C 0 T 1 − α μ T / C 2 , which tends to one by (J2) and leaves the limit unchanged. The paths of fractional Brownian motion with the Hurst parameter H H H satisfy the bound in Theorem 5.2 for every h < H h<H h < H [31 ] . In the limit, the intensity in units of its mean is thus as rough as fractional Brownian motion with the Hurst parameter H = α − 1 / 2 H=\alpha-1/2 H = α − 1/2 .
The proof of Theorem 5.2 passes to the limit in the representation of the intensity λ T \lambda^{\mathcal T} λ T in three steps. Let
Z t T = ( 1 − ∥ K T ∥ T μ T ) 1 / 2 M T t T , t ∈ [ 0 , 1 ] , Z^{\mathcal T}_t=\left(\frac{1-\lVert K_{\mathcal T}\rVert}{\mathcal T\mu_{\mathcal T}}\right)^{1/2}M^{\mathcal T}_{\mathcal Tt},\qquad t\in[0,1], Z t T = ( T μ T 1 − ∥ K T ∥ ) 1/2 M T t T , t ∈ [ 0 , 1 ] ,
denote the rescaled deviation of the number of actions from its integrated intensity. Since N T N^{\mathcal T} N T jumps by one, the deviation M T M^{\mathcal T} M T satisfies
⟨ M T ⟩ t = ∫ 0 t λ s T d s , [ M T , M T ] t = N t T , \langle M^{\mathcal T}\rangle_t=\int_0^t\lambda^{\mathcal T}_s\,ds,\qquad[M^{\mathcal T},M^{\mathcal T}]_t=N^{\mathcal T}_t, ⟨ M T ⟩ t = ∫ 0 t λ s T d s , [ M T , M T ] t = N t T ,
where ⟨ M T ⟩ \langle M^{\mathcal T}\rangle ⟨ M T ⟩ is the process that makes ( M T ) 2 − ⟨ M T ⟩ (M^{\mathcal T})^2-\langle M^{\mathcal T}\rangle ( M T ) 2 − ⟨ M T ⟩ a martingale, and [ M T , M T ] [M^{\mathcal T},M^{\mathcal T}] [ M T , M T ] is the sum of the squared jumps of M T M^{\mathcal T} M T . In the units of X T X^{\mathcal T} X T , these two identities read
⟨ Z T ⟩ t = X t T − ( 1 − ∥ K T ∥ T μ T ) 1 / 2 Z t T , [ Z T , Z T ] t = X t T . \langle Z^{\mathcal T}\rangle_t=X^{\mathcal T}_t-\left(\frac{1-\lVert K_{\mathcal T}\rVert}{\mathcal T\mu_{\mathcal T}}\right)^{1/2}Z^{\mathcal T}_t,\qquad[Z^{\mathcal T},Z^{\mathcal T}]_t=X^{\mathcal T}_t. ⟨ Z T ⟩ t = X t T − ( T μ T 1 − ∥ K T ∥ ) 1/2 Z t T , [ Z T , Z T ] t = X t T .
In the first step, we obtain E [ sup t ≤ 1 ∣ Z t T ∣ 2 ] ≤ 4 E [ ⟨ Z T ⟩ 1 ] \mathbb E[\sup_{t\leq1}|Z^{\mathcal T}_t|^2]\leq4\mathbb E[\langle Z^{\mathcal T}\rangle_1] E [ sup t ≤ 1 ∣ Z t T ∣ 2 ] ≤ 4 E [⟨ Z T ⟩ 1 ] from Doob’s inequality. The last term in ⟨ Z T ⟩ \langle Z^{\mathcal T}\rangle ⟨ Z T ⟩ thus vanishes as T → ∞ \mathcal T\to\infty T → ∞ . Every limit Z Z Z is then a continuous martingale with the quadratic variation X X X . In the second step, we insert the representation of λ T \lambda^{\mathcal T} λ T into X T X^{\mathcal T} X T and let ρ T → G \rho^{\mathcal T}\to\mathcal G ρ T → G , which yields the equation in Theorem 5.2 . Its Brownian motion B B B comes from Z Z Z , as a continuous martingale with the quadratic variation X X X is a Brownian motion run at the time X X X ,
Z t = B X t , t ∈ [ 0 , 1 ] . Z_t= B_{X_t},\qquad t\in[0,1]. Z t = B X t , t ∈ [ 0 , 1 ] .
In the third step, the bound on the increments of X ′ \mathcal{X}' X ′ follows from the singularity of G G G at zero. Of these steps, only the two identities for M T M^T M T depend on the desks acting one at a time. In § 5.2 we derive their analogues for the size of the updates of our market.
The size of the updates as a number of actions
The proof of Theorem 5.2 depends on single actions only through two identities for the deviation M T M^T M T (§ 5.1 ). The first, ⟨ M T ⟩ t = ∫ 0 t λ s T d s \langle M^T\rangle_t=\int_0^t\lambda_s^T\,\mathrm{d}s ⟨ M T ⟩ t = ∫ 0 t λ s T d s , states that the conditional variance of the number of actions equals its intensity. For our market, we therefore compute the conditional variance of the size of the updates. As for κ l \kappa_l κ l in Proposition 4.5 , we freeze the contribution c m i ∣ Δ Z i ∣ c m_i|\Delta Z_i| c m i ∣Δ Z i ∣ of each acting desk i i i to the size of the updates (7) at its average. Now, let
ν l = 1 κ l w l ∑ i : k ( i ) = l ( c m i ∣ Δ Z i ∣ ) 2 \nu_l=\frac{1}{\kappa_l w_l}\sum_{i:k(i)=l}(c m_i|\Delta Z_i|)^2 ν l = κ l w l 1 i : k ( i ) = l ∑ ( c m i ∣Δ Z i ∣ ) 2
denote the contribution of one desk of cluster l l l , averaged with the weights c m i ∣ Δ Z i ∣ / ( κ l w l ) c m_i|\Delta Z_i|/(\kappa_lw_l) c m i ∣Δ Z i ∣/ ( κ l w l ) . Since the desks act independently given F n \mathcal{F}_n F n (§ 2.5 ), we obtain as in Lemma 4.2
E [ I n ∣ F n ] = ∑ l = 1 K κ l w l p l ( n ) , Var [ I n ∣ F n ] = ∑ l = 1 K ν l ( 1 − p l ( n ) ) κ l w l p l ( n ) . \mathbb{E}[I_n\mid\mathcal{F}_n]=\sum_{l=1}^{K}\kappa_l w_l p_l(n),\qquad\operatorname{Var}[I_n\mid\mathcal{F}_n]=\sum_{l=1}^{K}\nu_l(1-p_l(n))\kappa_l w_l p_l(n). E [ I n ∣ F n ] = l = 1 ∑ K κ l w l p l ( n ) , Var [ I n ∣ F n ] = l = 1 ∑ K ν l ( 1 − p l ( n )) κ l w l p l ( n ) .
If the clusters share the probability of acting, p l ( n ) = p ( n ) p_l(n)=p(n) p l ( n ) = p ( n ) , the conditional variance becomes
Var [ I n ∣ F n ] = ν ˉ ( 1 − p ( n ) ) E [ I n ∣ F n ] , ν ˉ = ∑ l ν l κ l w l ∑ l κ l w l . \operatorname{Var}[I_n\mid\mathcal{F}_n]=\bar{\nu}(1-p(n))\mathbb{E}[I_n\mid\mathcal{F}_n],\qquad\bar{\nu}=\frac{\sum_l\nu_l\kappa_lw_l}{\sum_l\kappa_lw_l}. Var [ I n ∣ F n ] = ν ˉ ( 1 − p ( n )) E [ I n ∣ F n ] , ν ˉ = ∑ l κ l w l ∑ l ν l κ l w l .
In units of ν ˉ \bar{\nu} ν ˉ , the size of the updates thus has the conditional variance of a number of actions, up to the factor 1 − p ( n ) 1-p(n) 1 − p ( n ) . If the 2 N 2N 2 N desks of cluster l l l contribute equally, then ν l = κ l w l / ( 2 N ) \nu_l=\kappa_lw_l/(2N) ν l = κ l w l / ( 2 N ) , such that
ν ˉ = 1 2 N ∑ l ( κ l w l ) 2 ∑ l κ l w l \bar{\nu}=\frac{1}{2N}\frac{\sum_l(\kappa_lw_l)^2}{\sum_l\kappa_lw_l} ν ˉ = 2 N 1 ∑ l κ l w l ∑ l ( κ l w l ) 2
decreases as 1 / N 1/N 1/ N .
In units of ν ˉ \bar{\nu} ν ˉ , the size of the updates accumulated up to the time t = n γ t=n\gamma t = nγ thus takes the role of the number of actions. By Proposition 4.5 , the size of the updates has the base rate p 0 κ ˉ p_0\bar{\kappa} p 0 κ ˉ per step. By Proposition 4.9 , the kernel per unit of time is ϕ ˉ = ∥ ϕ ∥ K \bar{\phi}=\|\phi\|\mathcal{K} ϕ ˉ = ∥ ϕ ∥ K , which tends to ∥ ϕ ∥ K ∞ \|\phi\|\mathcal{K}_\infty ∥ ϕ ∥ K ∞ for many desks (Proposition 4.11 ). The quantities in Theorem 5.2 thus correspond to ours as
N t T ⟷ 1 ν ˉ ∑ n ≤ t / γ I n , μ T ⟷ p 0 κ ˉ ν ˉ γ , ∥ K T ∥ K ⟷ ∥ ϕ ∥ K ∞ . N_t^T\longleftrightarrow\frac{1}{\bar{\nu}}\sum_{n\le t/\gamma}I_n,\qquad\mu_T\longleftrightarrow\frac{p_0\bar{\kappa}}{\bar{\nu}\gamma},\qquad\|\mathcal{K}_T\|_{\mathcal{K}}\longleftrightarrow\|\phi\|\mathcal{K}_\infty. N t T ⟷ ν ˉ 1 n ≤ t / γ ∑ I n , μ T ⟷ ν ˉ γ p 0 κ ˉ , ∥ K T ∥ K ⟷ ∥ ϕ ∥ K ∞ .
Condition (J1) holds by Proposition 4.11 for α ∈ ( 1 / 2 , 1 ) \alpha\in(1/2,1) α ∈ ( 1/2 , 1 ) . With the time scale T \mathcal{T} T (§ 5.1 ), we have T α ( 1 − ∥ ϕ ∥ ) = ∥ ϕ ∥ C 0 → C 0 \mathcal{T}^{\alpha}(1-\|\phi\|)=\|\phi\|C_0\to C_0 T α ( 1 − ∥ ϕ ∥ ) = ∥ ϕ ∥ C 0 → C 0 , such that the first condition in (J2) holds with C 1 = 1 C_1=1 C 1 = 1 . Inserting T − α = ( 1 − ∥ ϕ ∥ ) / ( ∥ ϕ ∥ C 0 ) \mathcal{T}^{-\alpha}=(1-\|\phi\|)/(\|\phi\|C_0) T − α = ( 1 − ∥ ϕ ∥ ) / ( ∥ ϕ ∥ C 0 ) , we write the second condition in (J2) as
T ( 1 − ∥ ϕ ∥ ) p 0 ∥ ϕ ∥ γ ν ˉ / κ ˉ → C 2 . \frac{\mathcal{T}(1-\|\phi\|)p_0}{\|\phi\|\gamma\bar{\nu}/\bar{\kappa}}\to C_2. ∥ ϕ ∥ γ ν ˉ / κ ˉ T ( 1 − ∥ ϕ ∥ ) p 0 → C 2 .
By Proposition 4.8 , we have p 0 ≤ p ˉ ( 1 − ∥ ϕ ∥ ) p_0\le\bar{p}(1-\|\phi\|) p 0 ≤ p ˉ ( 1 − ∥ ϕ ∥ ) , such that
ν ˉ κ ˉ ≤ T ( 1 − ∥ ϕ ∥ ) 2 p ˉ ∥ ϕ ∥ C 2 ( 1 + o ( 1 ) ) = ∥ ϕ ∥ C 0 2 p ˉ C 2 T 1 − 2 α ( 1 + o ( 1 ) ) → 0 for α > 1 / 2. \frac{\bar{\nu}}{\bar{\kappa}}\le\frac{\mathcal{T}(1-\|\phi\|)^2\bar{p}}{\|\phi\|C_2}(1+o(1))=\frac{\|\phi\|C_0^2\bar{p}}{C_2}\mathcal{T}^{1-2\alpha}(1+o(1))\to0\qquad\text{for }\alpha>1/2. κ ˉ ν ˉ ≤ ∥ ϕ ∥ C 2 T ( 1 − ∥ ϕ ∥ ) 2 p ˉ ( 1 + o ( 1 )) = C 2 ∥ ϕ ∥ C 0 2 p ˉ T 1 − 2 α ( 1 + o ( 1 )) → 0 for α > 1/2.
Since ν ˉ / κ ˉ \bar{\nu}/\bar{\kappa} ν ˉ / κ ˉ is proportional to 1 / N 1/N 1/ N , the ratio γ / N \gamma/N γ / N of the step length to the number of desks per cluster and side must vanish at least as fast as T 1 − 2 α \mathcal{T}^{1-2\alpha} T 1 − 2 α . A market observed over longer times must therefore hold more desks or act in shorter steps.
The limit of the size of the updates
The correspondence of § 5.2 expresses the conditions (J1) and (J2) of Theorem 5.2 in the quantities of our market. Yet the proof of Theorem 5.2 also uses the two identities for the deviation M T \mathcal{M}^{T} M T (§ 5.1 ), which depend on the desks acting one at a time. In our market, every cluster instead enters the conditional variance of the size of the updates with its own probability of acting p l ( n ) p_l(n) p l ( n ) and the factor 1 − p l ( n ) 1-p_l(n) 1 − p l ( n ) (§ 5.2 ). Moreover, the size of the updates in units of ν ˉ \bar{\nu} ν ˉ jumps by I n / ν ˉ I_n/\bar{\nu} I n / ν ˉ rather than by one. In the limit of many desks with vanishing mean activity p 0 / ( 1 − ∥ ϕ ∥ ) p_0/(1-\lVert\phi\rVert) p 0 / ( 1 − ∥ ϕ ∥) , both identities hold up to vanishing errors. We also let the horizons extend beyond T \mathcal{T} T and the steps shrink relative to T \mathcal{T} T , such that the kernel keeps its power law up to T \mathcal{T} T and sums over steps become integrals. Inserting the correspondence into X T \mathcal{X}^{T} X T (§ 5.1 ), we obtain
X t T = 1 − ∥ ϕ ∥ T ν ˉ γ p 0 κ ˉ 1 ν ˉ ∑ n ≤ T t / γ I n = γ T ∑ n ≤ T t / γ I n I ˉ , t ∈ [ 0 , 1 ] , \mathcal{X}^{T}_{t}=\frac{1-\lVert\phi\rVert}{\mathcal{T}}\frac{\bar{\nu}\gamma}{p_0\bar{\kappa}}\frac{1}{\bar{\nu}}\sum_{n\leq\mathcal{T}t/\gamma}I_n=\frac{\gamma}{\mathcal{T}}\sum_{n\leq\mathcal{T}t/\gamma}\frac{I_n}{\bar{I}},\qquad t\in[0,1], X t T = T 1 − ∥ ϕ ∥ p 0 κ ˉ ν ˉ γ ν ˉ 1 n ≤ T t / γ ∑ I n = T γ n ≤ T t / γ ∑ I ˉ I n , t ∈ [ 0 , 1 ] ,
where I ˉ = p 0 κ ˉ / ( 1 − ∥ ϕ ∥ ) \bar{I}=p_0\bar{\kappa}/(1-\lVert\phi\rVert) I ˉ = p 0 κ ˉ / ( 1 − ∥ ϕ ∥) is the limit of E [ I n ] \mathbb{E}[I_n] E [ I n ] (Proposition 4.8 ). With both identities, the rescaled size of the updates X T \mathcal{X}^{T} X T then has the same limit as the rescaled number of actions in Theorem 5.2 .
Theorem 5.3 (Size of the updates close to the threshold). For every T > 0 \mathcal{T}>0 T > 0 , consider a market under Assumptions 3.2 and 3.3 with γ < τ 1 \gamma<\tau_1 γ < τ 1 , in which the 2 N 2N 2 N desks of each cluster make equal frozen contributions (§ 5.2 ) and the sizes of the updates per unit activity are proportional to κ ( τ l ) \kappa(\tau_l) κ ( τ l ) , with κ \kappa κ as in Proposition 4.11 and independent of T \mathcal{T} T . Let T α ( 1 − ∥ ϕ ∥ ) = ∥ ϕ ∥ C 0 \mathcal{T}^{\alpha}(1-\lVert\phi\rVert)=\lVert\phi\rVert C_0 T α ( 1 − ∥ ϕ ∥) = ∥ ϕ ∥ C 0 with α ∈ ( 1 / 2 , 1 ) \alpha\in(1/2,1) α ∈ ( 1/2 , 1 ) , and assume that, as T → ∞ \mathcal{T}\to\infty T → ∞ ,
(R1) Ω → 1 + \Omega\to1^{+} Ω → 1 + , τ K / T → ∞ \tau_K/\mathcal{T}\to\infty τ K / T → ∞ and γ / T → 0 \gamma/\mathcal{T}\to0 γ / T → 0 , while τ 1 \tau_1 τ 1 stays fixed;
(R2) p 0 / ( 1 − ∥ ϕ ∥ ) → 0 p_0/(1-\lVert\phi\rVert)\to0 p 0 / ( 1 − ∥ ϕ ∥) → 0 and T ( 1 − ∥ ϕ ∥ ) p 0 / ( ∥ ϕ ∥ γ ν ˉ / κ ˉ ) → C 2 \mathcal{T}(1-\lVert\phi\rVert)p_0/(\lVert\phi\rVert\gamma\bar{\nu}/\bar{\kappa})\to C_2 T ( 1 − ∥ ϕ ∥) p 0 / (∥ ϕ ∥ γ ν ˉ / κ ˉ ) → C 2 for some C 2 > 0 C_2>0 C 2 > 0 .
Then every sequence T k → ∞ \mathcal{T}_k\to\infty T k → ∞ has a subsequence along which X T k \mathcal{X}^{\mathcal{T}_k} X T k converges in law. Every such limit X \mathcal{X} X is differentiable, and there is a Brownian motion B B B such that
X t ′ = ∫ 0 t G ( s ) d s + 1 C 2 ∫ 0 t G ( t − s ) X s ′ d B s , t ∈ [ 0 , 1 ] , \mathcal{X}'_t=\int_0^t G(s)\,\mathrm{d}s+\frac{1}{\sqrt{C_2}}\int_0^t G(t-s)\sqrt{\mathcal{X}'_s}\,\mathrm{d}B_s,\qquad t\in[0,1], X t ′ = ∫ 0 t G ( s ) d s + C 2 1 ∫ 0 t G ( t − s ) X s ′ d B s , t ∈ [ 0 , 1 ] , where G G G has the Laplace transform 1 / ( 1 + z α ) 1/(1+z^\alpha) 1/ ( 1 + z α ) . Moreover, for every h < α − 1 / 2 h<\alpha-1/2 h < α − 1/2 ,
sup 0 ≤ s < t ≤ 1 ∣ X t ′ − X s ′ ∣ ( t − s ) h < ∞ almost surely. \sup_{0\leq s<t\leq1}\frac{\lvert\mathcal{X}'_t-\mathcal{X}'_s\rvert}{(t-s)^h}<\infty\qquad\text{almost surely.} 0 ≤ s < t ≤ 1 sup ( t − s ) h ∣ X t ′ − X s ′ ∣ < ∞ almost surely.
In the limit, the expected size of the updates in units of I ˉ \bar{I} I ˉ is thus as rough as fractional Brownian motion with the Hurst parameter $H=\alpha-1/2.
The first identity sums the conditional variance over all steps up to the time T t \mathcal{T}t T t . By § 5.2 , each cluster l l l enters this sum through ∑ r p l ( r ) ( 1 − p l ( r ) ) \sum_r p_l(r)(1-p_l(r)) ∑ r p l ( r ) ( 1 − p l ( r )) . At a given step, clusters with different horizons have different probabilities of acting, as their moving averages U l U_l U l weigh the past size of the updates differently (5). Summed over all later steps, however, these weights equal one for every cluster (§ 4.1 ). We therefore compare the probabilities of acting summed over steps.
Lemma 5.4 (Equal total response of the clusters). Under Assumption 3.3 (a), for every cluster l l l and every step n ∈ N n \in\mathbb{N} n ∈ N ,
∑ r = 0 n p l ( r ) = ( n + 1 ) p 0 + β ∑ m = 0 n − 1 I m − β ( T l − 1 ) U l ( n ) . \sum_{r=0}^{n} p_l(r) = (n+1)p_0 + \beta\sum_{m=0}^{n-1} I_m - \beta(T_l-1)U_l(n). r = 0 ∑ n p l ( r ) = ( n + 1 ) p 0 + β m = 0 ∑ n − 1 I m − β ( T l − 1 ) U l ( n ) . Every cluster thus responds to each past update with the same total, and the clusters differ only in the part of this response that falls after the step n n n .
Proof. By Assumption 3.3 (a), p l ( r ) = p 0 + β U l ( r ) p_l(r)=p_0+\beta U_l(r) p l ( r ) = p 0 + β U l ( r ) with U l ( r ) = ∑ m = 0 r − 1 ω l ( r − m ) I m U_l(r)=\sum_{m=0}^{r-1}\omega_l(r-m)I_m U l ( r ) = ∑ m = 0 r − 1 ω l ( r − m ) I m (proof of Proposition 4.3 ). Exchanging the sums and summing the geometric weights ω l ( s ) = ( 1 − 1 / T l ) s − 1 / T l \omega_l(s)=(1-1/T_l)^{s-1}/T_l ω l ( s ) = ( 1 − 1/ T l ) s − 1 / T l , we obtain
∑ r = 0 n U l ( r ) = ∑ m = 0 n − 1 I m ∑ s = 1 n − m ω l ( s ) = ∑ m = 0 n − 1 I m ( 1 − ( 1 − 1 T l ) n − m ) . \sum_{r=0}^{n} U_l(r)=\sum_{m=0}^{n-1}I_m\sum_{s=1}^{n-m}\omega_l(s)=\sum_{m=0}^{n-1}I_m\left(1-\left(1-\frac{1}{T_l}\right)^{n-m}\right). r = 0 ∑ n U l ( r ) = m = 0 ∑ n − 1 I m s = 1 ∑ n − m ω l ( s ) = m = 0 ∑ n − 1 I m ( 1 − ( 1 − T l 1 ) n − m ) . Since ( 1 − 1 / T l ) s = ( T l − 1 ) ω l ( s ) (1-1/T_l)^s=(T_l-1)\omega_l(s) ( 1 − 1/ T l ) s = ( T l − 1 ) ω l ( s ) for every s ∈ N ∖ { 0 } s\in\mathbb{N}\setminus\{0\} s ∈ N ∖ { 0 } , the last term satisfies
∑ m = 0 n − 1 I m ( 1 − 1 T l ) n − m = ( T l − 1 ) ∑ m = 0 n − 1 ω l ( n − m ) I m = ( T l − 1 ) U l ( n ) . \sum_{m=0}^{n-1}I_m\left(1-\frac{1}{T_l}\right)^{n-m}=(T_l-1)\sum_{m=0}^{n-1}\omega_l(n-m)I_m=(T_l-1)U_l(n). m = 0 ∑ n − 1 I m ( 1 − T l 1 ) n − m = ( T l − 1 ) m = 0 ∑ n − 1 ω l ( n − m ) I m = ( T l − 1 ) U l ( n ) . Multiplying by β \beta β and adding ( n + 1 ) p 0 (n+1)p_0 ( n + 1 ) p 0 , we obtain the claim. □
Inserting Lemma 5.4 into the summed conditional variance, we obtain ν ˉ \bar{\nu} ν ˉ times the summed conditional mean up to two errors.
Lemma 5.5 (Conditional variance summed over steps). Under Assumption 3.3 (a) and with frozen contributions of the desks (§ 5.2 ), for every step n ∈ N n\in\mathbb{N} n ∈ N ,
∑ r = 0 n Var [ I r ∣ F r ] = ν ˉ ∑ r = 0 n E [ I r ∣ F r ] − β ∑ l = 1 K ( ν l − ν ˉ ) κ l w l ( T l − 1 ) U l ( n ) − ∑ l = 1 K ν l κ l w l ∑ r = 0 n p l ( r ) 2 . \sum_{r=0}^{n}\operatorname{Var}[I_r\mid\mathcal{F}_r]
=
\bar{\nu}\sum_{r=0}^{n}\mathbb{E}[I_r\mid\mathcal{F}_r]
-\beta\sum_{l=1}^{K}(\nu_l-\bar{\nu})\kappa_l w_l(T_l-1)U_l(n)
-\sum_{l=1}^{K}\nu_l\kappa_l w_l\sum_{r=0}^{n}p_l(r)^2. r = 0 ∑ n Var [ I r ∣ F r ] = ν ˉ r = 0 ∑ n E [ I r ∣ F r ] − β l = 1 ∑ K ( ν l − ν ˉ ) κ l w l ( T l − 1 ) U l ( n ) − l = 1 ∑ K ν l κ l w l r = 0 ∑ n p l ( r ) 2 . By the proof of Lemma 5.4 and Assumption 3.2 , the part of the response of cluster l l l after the step n n n is
β ( T l − 1 ) U l ( n ) = β ∑ m = 0 n − 1 ( 1 − γ τ l ) n − m I m ≤ β ∑ m = 0 n − 1 e − ( n − m ) γ / τ l I m . \beta(T_l-1)U_l(n)=\beta\sum_{m=0}^{n-1}\left(1-\frac{\gamma}{\tau_l}\right)^{n-m}I_m\le\beta\sum_{m=0}^{n-1}e^{-(n-m)\gamma/\tau_l}I_m. β ( T l − 1 ) U l ( n ) = β m = 0 ∑ n − 1 ( 1 − τ l γ ) n − m I m ≤ β m = 0 ∑ n − 1 e − ( n − m ) γ / τ l I m . As p l ( r ) 2 ≤ p l ( r ) max l ′ , r ′ p l ′ ( r ′ ) p_l(r)^2\le p_l(r)\max_{l',r'}p_{l'}(r') p l ( r ) 2 ≤ p l ( r ) max l ′ , r ′ p l ′ ( r ′ ) , the second error satisfies
∑ l = 1 K ν l κ l w l ∑ r = 0 n p l ( r ) 2 ≤ max l ν l ν ˉ max l , r ≤ n p l ( r ) ν ˉ ∑ r = 0 n E [ I r ∣ F r ] . \sum_{l=1}^{K}\nu_l\kappa_l w_l\sum_{r=0}^{n}p_l(r)^2
\le
\max_l\frac{\nu_l}{\bar{\nu}}\max_{l,r\le n}p_l(r)\,\bar{\nu}\sum_{r=0}^{n}\mathbb{E}[I_r\mid\mathcal{F}_r]. l = 1 ∑ K ν l κ l w l r = 0 ∑ n p l ( r ) 2 ≤ l max ν ˉ ν l l , r ≤ n max p l ( r ) ν ˉ r = 0 ∑ n E [ I r ∣ F r ] . Proof. By § 5.2 , the conditional variance and the conditional mean of the size of the updates are ∑ l ν l κ l w l p l ( r ) ( 1 − p l ( r ) ) \sum_l\nu_l\kappa_lw_lp_l(r)(1-p_l(r)) ∑ l ν l κ l w l p l ( r ) ( 1 − p l ( r )) and ∑ l κ l w l p l ( r ) \sum_l\kappa_lw_lp_l(r) ∑ l κ l w l p l ( r ) . Summing over r r r and subtracting ν ˉ \bar{\nu} ν ˉ times the second sum from the first, we obtain
∑ r = 0 n Var [ I r ∣ F r ] − ν ˉ ∑ r = 0 n E [ I r ∣ F r ] = ∑ l = 1 K ( ν l − ν ˉ ) κ l w l ∑ r = 0 n p l ( r ) − ∑ l = 1 K ν l κ l w l ∑ r = 0 n p l ( r ) 2 . \sum_{r=0}^{n}\operatorname{Var}[I_r\mid\mathcal{F}_r]
-\bar{\nu}\sum_{r=0}^{n}\mathbb{E}[I_r\mid\mathcal{F}_r]
=
\sum_{l=1}^{K}(\nu_l-\bar{\nu})\kappa_lw_l\sum_{r=0}^{n}p_l(r)
-\sum_{l=1}^{K}\nu_l\kappa_lw_l\sum_{r=0}^{n}p_l(r)^2. r = 0 ∑ n Var [ I r ∣ F r ] − ν ˉ r = 0 ∑ n E [ I r ∣ F r ] = l = 1 ∑ K ( ν l − ν ˉ ) κ l w l r = 0 ∑ n p l ( r ) − l = 1 ∑ K ν l κ l w l r = 0 ∑ n p l ( r ) 2 .
By the definition of ν ˉ \bar{\nu} ν ˉ , the weights ( ν l − ν ˉ ) κ l w l (\nu_l-\bar{\nu})\kappa_lw_l ( ν l − ν ˉ ) κ l w l sum to zero over l l l . When we insert Lemma 5.4 into the first sum on the right, the part ( n + 1 ) p 0 + β ∑ m < n I m (n+1)p_0+\beta\sum_{m<n}I_m ( n + 1 ) p 0 + β ∑ m < n I m therefore cancels, and only − β ( T l − 1 ) U l ( n ) -\beta(T_l-1)U_l(n) − β ( T l − 1 ) U l ( n ) remains.
Taking expectations in the second error of Lemma 5.5 and inserting U l U_l U l from the proof of Proposition 4.3 , we obtain
E [ ∑ l = 1 K ν l κ l w l ∑ r = 0 n p l ( r ) 2 ] = ∑ l = 1 K ν l κ l w l ∑ r = 0 n E [ ( p 0 + β ∑ m = 0 r − 1 ω l ( r − m ) I m ) 2 ] . \mathbb{E}\left[\sum_{l=1}^{K}\nu_l\kappa_lw_l\sum_{r=0}^{n}p_l(r)^2\right]
=\sum_{l=1}^{K}\nu_l\kappa_lw_l\sum_{r=0}^{n}\mathbb{E}\left[\left(p_0+\beta\sum_{m=0}^{r-1}\omega_l(r-m)I_m\right)^2\right]. E [ l = 1 ∑ K ν l κ l w l r = 0 ∑ n p l ( r ) 2 ] = l = 1 ∑ K ν l κ l w l r = 0 ∑ n E ( p 0 + β m = 0 ∑ r − 1 ω l ( r − m ) I m ) 2 .
Since the deviations I ~ m = I m − E [ I m ∣ F m ] \widetilde{I}_m=I_m-\mathbb{E}[I_m\mid\mathcal{F}_m] I m = I m − E [ I m ∣ F m ] are uncorrelated, we obtain from the representation of the expected size of the updates (§ 5.1 )
E [ I n ∣ F n ] = p 0 κ ˉ ( 1 + ∑ s = 1 n R ( s ) ) + ∑ m = 0 n − 1 R ( n − m ) I ~ m , \mathbb{E}[I_n\mid\mathcal{F}_n]=p_0\bar{\kappa}\left(1+\sum_{s=1}^{n}R(s)\right)+\sum_{m=0}^{n-1}R(n-m)\widetilde{I}_m, E [ I n ∣ F n ] = p 0 κ ˉ ( 1 + s = 1 ∑ n R ( s ) ) + m = 0 ∑ n − 1 R ( n − m ) I m ,
Var ( E [ I n ∣ F n ] ) = ∑ m = 0 n − 1 R ( n − m ) 2 E [ I ~ m 2 ] ≤ max m < n E [ I ~ m 2 ] ∑ s ∈ N ∖ { 0 } R ( s ) 2 . \operatorname{Var}\left(\mathbb{E}[I_n\mid\mathcal{F}_n]\right)=\sum_{m=0}^{n-1}R(n-m)^2\mathbb{E}[\widetilde{I}_m^2]\leq\max_{m<n}\mathbb{E}[\widetilde{I}_m^2]\sum_{s\in\mathbb{N}\setminus\{0\}}R(s)^2. Var ( E [ I n ∣ F n ] ) = m = 0 ∑ n − 1 R ( n − m ) 2 E [ I m 2 ] ≤ m < n max E [ I m 2 ] s ∈ N ∖ { 0 } ∑ R ( s ) 2 .
Lemma 5.6 (Square-summable response). Under the assumptions of Theorem 5.3 , there is a constant C R C_R C R , depending only on α \alpha α , τ 1 \tau_1 τ 1 , C 0 C_0 C 0 and inf κ / sup κ \inf\kappa/\sup\kappa inf κ / sup κ , such that for every sufficiently large T \mathcal{T} T
∑ s ∈ N ∖ { 0 } R ( s ) 2 ≤ C R γ ( 1 − ∥ ϕ ∥ ) 2 T . \sum_{s\in\mathbb{N}\setminus\{0\}}R(s)^2\leq C_R\frac{\gamma}{(1-\|\phi\|)^2\mathcal{T}}. s ∈ N ∖ { 0 } ∑ R ( s ) 2 ≤ C R ( 1 − ∥ ϕ ∥ ) 2 T γ .
By Proposition 4.8 , ∑ s R ( s ) = ∥ ϕ ∥ / ( 1 − ∥ ϕ ∥ ) \sum_sR(s)=\|\phi\|/(1-\|\phi\|) ∑ s R ( s ) = ∥ ϕ ∥/ ( 1 − ∥ ϕ ∥ ) , such that the lemma reads
∑ s ∈ N ∖ { 0 } R ( s ) 2 ≤ C R ∥ ϕ ∥ 2 γ T ( ∑ s ∈ N ∖ { 0 } R ( s ) ) 2 . \sum_{s\in\mathbb{N}\setminus\{0\}}R(s)^2\leq\frac{C_R}{\|\phi\|^2}\frac{\gamma}{\mathcal{T}}\left(\sum_{s\in\mathbb{N}\setminus\{0\}}R(s)\right)^2. s ∈ N ∖ { 0 } ∑ R ( s ) 2 ≤ ∥ ϕ ∥ 2 C R T γ s ∈ N ∖ { 0 } ∑ R ( s ) 2 .
Proof. By Proposition 4.8 , ∑ s R ( s ) = ∥ ϕ ∥ / ( 1 − ∥ ϕ ∥ ) < ∞ \sum_sR(s)=\|\phi\|/(1-\|\phi\|)<\infty ∑ s R ( s ) = ∥ ϕ ∥/ ( 1 − ∥ ϕ ∥ ) < ∞ . Parseval’s identity thus applies to R ^ ( z ) = ∑ s R ( s ) z s \hat{R}(z)=\sum_sR(s)z^s R ^ ( z ) = ∑ s R ( s ) z s , and as R R R is real,
∑ s ∈ N ∖ { 0 } R ( s ) 2 = 1 2 π ∫ − π π ∣ R ^ ( e i θ ) ∣ 2 d θ = 1 π ∫ 0 π ∣ R ^ ( e i θ ) ∣ 2 d θ . \sum_{s\in\mathbb{N}\setminus\{0\}}R(s)^2=\frac{1}{2\pi}\int_{-\pi}^{\pi}\left|\hat{R}(e^{i\theta})\right|^2\,d\theta=\frac{1}{\pi}\int_{0}^{\pi}\left|\hat{R}(e^{i\theta})\right|^2\,d\theta. s ∈ N ∖ { 0 } ∑ R ( s ) 2 = 2 π 1 ∫ − π π R ^ ( e i θ ) 2 d θ = π 1 ∫ 0 π R ^ ( e i θ ) 2 d θ .
By Assumption 3.2 , ω l ( s ) = ( 1 − a l ) a l s − 1 \omega_l(s)=(1-a_l)a_l^{s-1} ω l ( s ) = ( 1 − a l ) a l s − 1 with a l = 1 − γ / τ l a_l=1-\gamma/\tau_l a l = 1 − γ / τ l . Using β κ l w l = ∥ ϕ ∥ ϖ l \beta\kappa_lw_l=\|\phi\|\varpi_l β κ l w l = ∥ ϕ ∥ ϖ l (Proposition 4.7 ) in the kernel (15) and R = ∑ j ϕ ∗ j R=\sum_j\phi^{*j} R = ∑ j ϕ ∗ j (Corollary 4.6 ), we obtain for ∣ z ∣ ≤ 1 |z|\leq1 ∣ z ∣ ≤ 1
ϕ ^ ( z ) = ∥ ϕ ∥ ∑ l = 1 K ϖ l ∑ s ∈ N ∖ { 0 } ( 1 − a l ) a l s − 1 z s = ∥ ϕ ∥ ∑ l = 1 K ϖ l ( 1 − a l ) z 1 − a l z , \hat{\phi}(z)=\|\phi\|\sum_{l=1}^{K}\varpi_l\sum_{s\in\mathbb{N}\setminus\{0\}}(1-a_l)a_l^{s-1}z^s
=\|\phi\|\sum_{l=1}^{K}\varpi_l\frac{(1-a_l)z}{1-a_lz}, ϕ ^ ( z ) = ∥ ϕ ∥ l = 1 ∑ K ϖ l s ∈ N ∖ { 0 } ∑ ( 1 − a l ) a l s − 1 z s = ∥ ϕ ∥ l = 1 ∑ K ϖ l 1 − a l z ( 1 − a l ) z ,
1 − ϕ ^ ( z ) = ( 1 − ∥ ϕ ∥ ) + ∥ ϕ ∥ ∑ l = 1 K ϖ l 1 − z 1 − a l z , 1-\hat{\phi}(z)=(1-\|\phi\|)+\|\phi\|\sum_{l=1}^{K}\varpi_l\frac{1-z}{1-a_lz}, 1 − ϕ ^ ( z ) = ( 1 − ∥ ϕ ∥ ) + ∥ ϕ ∥ l = 1 ∑ K ϖ l 1 − a l z 1 − z ,
R ^ ( z ) = ∑ j ∈ N ∖ { 0 } ϕ ^ ( z ) j = ϕ ^ ( z ) 1 − ϕ ^ ( z ) , \hat{R}(z)=\sum_{j\in\mathbb{N}\setminus\{0\}}\hat{\phi}(z)^j=\frac{\hat{\phi}(z)}{1-\hat{\phi}(z)}, R ^ ( z ) = j ∈ N ∖ { 0 } ∑ ϕ ^ ( z ) j = 1 − ϕ ^ ( z ) ϕ ^ ( z ) ,
where the last series converges as $|\hat{\phi}(z)|\leq|\phi|<1. Next, let x = 1 − cos θ x=1-\cos\theta x = 1 − cos θ and y ( θ ) = π γ / ( 2 θ ) y(\theta)=\pi\gamma/(2\theta) y ( θ ) = π γ / ( 2 θ ) . Every cluster with τ l ≥ y ( θ ) \tau_l\ge y(\theta) τ l ≥ y ( θ ) satisfies ( γ / τ l ) 2 ≤ 2 x (\gamma/\tau_l)^2\le2x ( γ / τ l ) 2 ≤ 2 x , as x ≥ 2 θ 2 / π 2 x\ge2\theta^2/\pi^2 x ≥ 2 θ 2 / π 2 for ∣ θ ∣ ≤ π |\theta|\le\pi ∣ θ ∣ ≤ π . Hence
Re ( 1 − ϕ ^ ( e i θ ) ) = ( 1 − ∥ ϕ ∥ ) + ∥ ϕ ∥ ∑ l = 1 K ϖ l ( 1 + a l ) x ( 1 − a l ) 2 + 2 a l x ≥ ( 1 − ∥ ϕ ∥ ) + ∥ ϕ ∥ ∑ l = 1 K ϖ l x ( γ / τ l ) 2 + 2 x ≥ ( 1 − ∥ ϕ ∥ ) + ∥ ϕ ∥ 4 ∑ l : τ l ≥ y ( θ ) ϖ l . \begin{aligned}
\operatorname{Re}\left(1-\widehat{\phi}(e^{i\theta})\right)=(1-\|\phi\|)+\|\phi\|\sum_{l=1}^{K}\varpi_l\frac{(1+a_l)x}{(1-a_l)^2+2a_lx} \\
&\ge(1-\|\phi\|)+\|\phi\|\sum_{l=1}^{K}\varpi_l\frac{x}{(\gamma/\tau_l)^2+2x}\ge(1-\|\phi\|)+\frac{\|\phi\|}{4}\sum_{l:\tau_l\ge y(\theta)}\varpi_l.
\end{aligned} Re ( 1 − ϕ ( e i θ ) ) = ( 1 − ∥ ϕ ∥ ) + ∥ ϕ ∥ l = 1 ∑ K ϖ l ( 1 − a l ) 2 + 2 a l x ( 1 + a l ) x ≥ ( 1 − ∥ ϕ ∥ ) + ∥ ϕ ∥ l = 1 ∑ K ϖ l ( γ / τ l ) 2 + 2 x x ≥ ( 1 − ∥ ϕ ∥ ) + 4 ∥ ϕ ∥ l : τ l ≥ y ( θ ) ∑ ϖ l .
By the proof of Proposition 4.11 , ϖ l = κ l τ l − α / ∑ l κ l τ l − α \varpi_l=\kappa_l\tau_l^{-\alpha}/\sum_l\kappa_l\tau_l^{-\alpha} ϖ l = κ l τ l − α / ∑ l κ l τ l − α . For τ 1 ≤ y ≤ τ K / 2 \tau_1\le y\le\tau_K/2 τ 1 ≤ y ≤ τ K /2 and Ω ≤ 2 \Omega\le2 Ω ≤ 2 , let l 0 l_0 l 0 be the first cluster with τ l 0 ≥ y \tau_{l_0}\ge y τ l 0 ≥ y , such that τ l 0 < Ω y \tau_{l_0}<\Omega y τ l 0 < Ω y . Summing the geometric series in τ l − α \tau_l^{-\alpha} τ l − α and using τ K = τ l 0 Ω K − l 0 \tau_K=\tau_{l_0}\Omega^{K-l_0} τ K = τ l 0 Ω K − l 0 , we obtain
∑ l : τ l ≥ y ϖ l ≥ inf κ sup κ ( τ l 0 τ 1 ) − α ( 1 − Ω − α ( K − l 0 + 1 ) ) ≥ inf κ sup κ ( Ω y τ K ) − α ( 1 − ( y τ K ) α ) ≥ c ϖ ( τ 1 y ) α , c ϖ = inf κ sup κ 2 − α ( 1 − 2 − α ) , \begin{aligned}
\sum_{l:\tau_l\ge y}\varpi_l\ge\frac{\inf\kappa}{\sup\kappa}\left(\frac{\tau_{l_0}}{\tau_1}\right)^{-\alpha}\left(1-\Omega^{-\alpha(K-l_0+1)}\right) \\
&\ge\frac{\inf\kappa}{\sup\kappa}\left(\frac{\Omega y}{\tau_K}\right)^{-\alpha}\left(1-\left(\frac{y}{\tau_K}\right)^\alpha\right)\ge c_\varpi\left(\frac{\tau_1}{y}\right)^\alpha,\qquad c_\varpi=\frac{\inf\kappa}{\sup\kappa}2^{-\alpha}(1-2^{-\alpha}),
\end{aligned} l : τ l ≥ y ∑ ϖ l ≥ sup κ inf κ ( τ 1 τ l 0 ) − α ( 1 − Ω − α ( K − l 0 + 1 ) ) ≥ sup κ inf κ ( τ K Ω y ) − α ( 1 − ( τ K y ) α ) ≥ c ϖ ( y τ 1 ) α , c ϖ = sup κ inf κ 2 − α ( 1 − 2 − α ) ,
while for y < τ 1 y<\tau_1 y < τ 1 the sum equals one. With θ K = π γ / τ K \theta_K=\pi\gamma/\tau_K θ K = π γ / τ K , we have y ( θ ) ≤ τ K / 2 y(\theta)\le\tau_K/2 y ( θ ) ≤ τ K /2 for θ ≥ θ K \theta\ge\theta_K θ ≥ θ K . Inserting τ 1 / y ( θ ) = 2 τ 1 θ / ( π γ ) \tau_1/y(\theta)=2\tau_1\theta/(\pi\gamma) τ 1 / y ( θ ) = 2 τ 1 θ / ( π γ ) , we obtain for θ K ≤ θ ≤ π \theta_K\le\theta\le\pi θ K ≤ θ ≤ π
∣ 1 − ϕ ^ ( e i θ ) ∣ ≥ ( 1 − ∥ ϕ ∥ ) + c min { 1 , ( τ 1 θ γ ) α } , c = ∥ ϕ ∥ c ϖ 4 ( 2 π ) α . \left|1-\widehat{\phi}(e^{i\theta})\right|\ge(1-\|\phi\|)+c\min\left\{1,\left(\frac{\tau_1\theta}{\gamma}\right)^\alpha\right\},\qquad c=\frac{\|\phi\|c_\varpi}{4}\left(\frac{2}{\pi}\right)^\alpha. 1 − ϕ ( e i θ ) ≥ ( 1 − ∥ ϕ ∥ ) + c min { 1 , ( γ τ 1 θ ) α } , c = 4 ∥ ϕ ∥ c ϖ ( π 2 ) α .
Lastly, we split the integral at θ K \theta_K θ K and at γ / τ 1 \gamma/\tau_1 γ / τ 1 , and use ∣ R ^ ∣ ≤ 1 / ( 1 − ∥ ϕ ∥ ) |\widehat{R}|\le1/(1-\|\phi\|) ∣ R ∣ ≤ 1/ ( 1 − ∥ ϕ ∥ ) , ∣ ϕ ^ ∣ ≤ 1 |\widehat{\phi}|\le1 ∣ ϕ ∣ ≤ 1 and ∣ 1 − ϕ ^ ∣ ≥ c |1-\widehat{\phi}|\ge c ∣1 − ϕ ∣ ≥ c on the three parts. Substituting θ = ( γ / τ 1 ) ( ( 1 − ∥ ϕ ∥ ) / c ) 1 / α u \theta=(\gamma/\tau_1)((1-\|\phi\|)/c)^{1/\alpha}u θ = ( γ / τ 1 ) (( 1 − ∥ ϕ ∥ ) / c ) 1/ α u in the second part and using ∑ s ϕ ( s ) 2 ≤ ϕ ( 1 ) ∥ ϕ ∥ ≤ γ / τ 1 \sum_s\phi(s)^2\le\phi(1)\|\phi\|\le\gamma/\tau_1 ∑ s ϕ ( s ) 2 ≤ ϕ ( 1 ) ∥ ϕ ∥ ≤ γ / τ 1 for the decreasing ϕ \phi ϕ in the third, we obtain with J α = ∫ 0 ∞ d u / ( 1 + u α ) 2 \mathcal{J}_\alpha=\int_0^\infty du/(1+u^\alpha)^2 J α = ∫ 0 ∞ d u / ( 1 + u α ) 2
∑ s R ( s ) 2 ≤ 1 π ∫ 0 θ K d θ ( 1 − ∥ ϕ ∥ ) 2 + 1 π ∫ 0 ∞ d θ ( ( 1 − ∥ ϕ ∥ ) + c ( τ 1 θ / γ ) α ) 2 + 1 π c 2 ∫ 0 π ∣ ϕ ^ ( e i θ ) ∣ 2 d θ ≤ γ τ K ( 1 − ∥ ϕ ∥ ) 2 + γ π τ 1 c − 1 / α ( 1 − ∥ ϕ ∥ ) 1 / α − 2 J α + γ c 2 τ 1 . \begin{aligned}
\sum_sR(s)^2\le\frac{1}{\pi}\int_0^{\theta_K}\frac{d\theta}{(1-\|\phi\|)^2}+\frac{1}{\pi}\int_0^\infty\frac{d\theta}{((1-\|\phi\|)+c(\tau_1\theta/\gamma)^\alpha)^2}+\frac{1}{\pi c^2}\int_0^\pi|\widehat{\phi}(e^{i\theta})|^2d\theta\\
&\le\frac{\gamma}{\tau_K(1-\|\phi\|)^2}+\frac{\gamma}{\pi\tau_1}c^{-1/\alpha}(1-\|\phi\|)^{1/\alpha-2}\mathcal{J}_\alpha+\frac{\gamma}{c^2\tau_1}.
\end{aligned} s ∑ R ( s ) 2 ≤ π 1 ∫ 0 θ K ( 1 − ∥ ϕ ∥ ) 2 d θ + π 1 ∫ 0 ∞ (( 1 − ∥ ϕ ∥ ) + c ( τ 1 θ / γ ) α ) 2 d θ + π c 2 1 ∫ 0 π ∣ ϕ ( e i θ ) ∣ 2 d θ ≤ τ K ( 1 − ∥ ϕ ∥ ) 2 γ + π τ 1 γ c − 1/ α ( 1 − ∥ ϕ ∥ ) 1/ α − 2 J α + c 2 τ 1 γ .
Inserting ( 1 − ∥ ϕ ∥ ) 1 / α = ( ∥ ϕ ∥ C 0 ) 1 / α / T (1-\|\phi\|)^{1/\alpha}=(\|\phi\|C_0)^{1/\alpha}/\mathcal{T} ( 1 − ∥ ϕ ∥ ) 1/ α = ( ∥ ϕ ∥ C 0 ) 1/ α / T , we obtain
∑ s R ( s ) 2 ≤ γ ( 1 − ∥ ϕ ∥ ) 2 T ( T τ K + J α π τ 1 ( ∥ ϕ ∥ C 0 c ) 1 / α + ( 1 − ∥ ϕ ∥ ) 2 T c 2 τ 1 ) , \sum_sR(s)^2\le\frac{\gamma}{(1-\|\phi\|)^2\mathcal{T}}\left(\frac{\mathcal{T}}{\tau_K}+\frac{\mathcal{J}_\alpha}{\pi\tau_1}\left(\frac{\|\phi\|C_0}{c}\right)^{1/\alpha}+\frac{(1-\|\phi\|)^2\mathcal{T}}{c^2\tau_1}\right), s ∑ R ( s ) 2 ≤ ( 1 − ∥ ϕ ∥ ) 2 T γ ( τ K T + π τ 1 J α ( c ∥ ϕ ∥ C 0 ) 1/ α + c 2 τ 1 ( 1 − ∥ ϕ ∥ ) 2 T ) ,
where J α \mathcal{J}_\alpha J α is finite because 2 α > 1 2\alpha>1 2 α > 1 . By (R1), T / τ K → 0 \mathcal{T}/\tau_K\to0 T / τ K → 0 , Ω ≤ 2 \Omega\le2 Ω ≤ 2 and τ K ≥ π τ 1 \tau_K\ge\pi\tau_1 τ K ≥ π τ 1 for large T \mathcal{T} T . Moreover, ( 1 − ∥ ϕ ∥ ) 2 T = ( ∥ ϕ ∥ C 0 ) 2 T 1 − 2 α → 0 (1-\|\phi\|)^2\mathcal{T}=(\|\phi\|C_0)^2\mathcal{T}^{1-2\alpha}\to0 ( 1 − ∥ ϕ ∥ ) 2 T = ( ∥ ϕ ∥ C 0 ) 2 T 1 − 2 α → 0 and ∥ ϕ ∥ → 1 \|\phi\|\to1 ∥ ϕ ∥ → 1 . The bracket thus converges to a limit that depends only on α \alpha α , τ 1 \tau_1 τ 1 , C 0 C_0 C 0 and inf κ / sup κ \inf\kappa/\sup\kappa inf κ / sup κ , and twice this limit is a valid C R C_R C R . □
Inserting Lemma 5.6 into the variance of the expected size of the updates E [ I n ∣ F n ] \mathbb{E}[I_n\mid\mathcal{F}_n] E [ I n ∣ F n ] , we obtain
Var ( E [ I n ∣ F n ] ) ≤ max m ≤ n E [ I ~ m 2 ] C R γ ( 1 − ∥ ϕ ∥ ) 2 T . \operatorname{Var}\left(\mathbb{E}[I_n\mid\mathcal{F}_n]\right)\le\max_{m\le n}\mathbb{E}[\widetilde{I}_m^2]C_R\frac{\gamma}{(1-\|\phi\|)^2\mathcal{T}}. Var ( E [ I n ∣ F n ] ) ≤ m ≤ n max E [ I m 2 ] C R ( 1 − ∥ ϕ ∥ ) 2 T γ .
With the conditional variance of § 5.2 and the mean E [ I m ] ≤ I ‾ \mathbb{E}[I_m]\le\overline{I} E [ I m ] ≤ I (Proposition 4.8 ), the second moment of each deviation I ~ m \widetilde{I}_m I m satisfies
E [ I ~ m 2 ] = E [ ∑ l = 1 K ν l ( 1 − p l ( m ) ) κ l w l p l ( m ) ] ≤ max l ν l E [ I m ] ≤ max l ν l ν ‾ ν ‾ I ‾ . \mathbb{E}[\widetilde{I}_m^2]=\mathbb{E}\left[\sum_{l=1}^{K}\nu_l(1-p_l(m))\kappa_lw_lp_l(m)\right]\le\max_l\nu_l\,\mathbb{E}[I_m]\le\max_l\frac{\nu_l}{\overline{\nu}}\,\overline{\nu}\,\overline{I}. E [ I m 2 ] = E [ l = 1 ∑ K ν l ( 1 − p l ( m )) κ l w l p l ( m ) ] ≤ l max ν l E [ I m ] ≤ l max ν ν l ν I .
Lemma 5.7 (Second moments). Under the assumptions of Theorem 5.3 , let q = p 0 / ( 1 − ∥ ϕ ∥ ) q = p_0/(1-\lVert\phi\rVert) q = p 0 / ( 1 − ∥ ϕ ∥) denote the mean activity of Proposition 4.8 , and let
C ν = 4 ( sup κ inf κ ) 3 , C λ = 1 + 2 C ν C R C 2 . C_{\nu}=4\left(\frac{\sup\kappa}{\inf\kappa}\right)^3,\qquad C_{\lambda}=1+\frac{2C_{\nu}C_R}{C_2}. C ν = 4 ( inf κ sup κ ) 3 , C λ = 1 + C 2 2 C ν C R . Then for every sufficiently large T T T , every cluster l l l and every step n n n ,
(i) ν l ≤ C ν ν ˉ , (ii) E [ E [ I n ∣ F n ] 2 ] ≤ C λ I ˉ 2 , (iii) E [ p l ( n ) 2 ] ≤ 3 ( 1 + C λ ) q 2 + 3 C ν q γ ν ˉ κ ˉ τ 1 . \begin{aligned}
\text{(i)}\quad&\nu_l\le C_{\nu}\bar{\nu}, \qquad
\text{(ii)}\quad\mathbb{E}\left[\mathbb{E}[I_n\mid\mathcal{F}_n]^2\right]\le C_{\lambda}\bar{I}^2,\\
\text{(iii)}\quad&\mathbb{E}[p_l(n)^2]\le3(1+C_{\lambda})q^2+3C_{\nu}q\frac{\gamma\bar{\nu}}{\bar{\kappa}\tau_1}.
\end{aligned} (i) (iii) ν l ≤ C ν ν ˉ , (ii) E [ E [ I n ∣ F n ] 2 ] ≤ C λ I ˉ 2 , E [ p l ( n ) 2 ] ≤ 3 ( 1 + C λ ) q 2 + 3 C ν q κ ˉ τ 1 γ ν ˉ .
By (R2), q → 0 q\to0 q → 0 , and by § 5.2 , ν ˉ / κ ˉ → 0 \bar{\nu}/\bar{\kappa}\to0 ν ˉ / κ ˉ → 0 . Hence E [ p l ( n ) 2 ] / q → 0 \mathbb{E}[p_l(n)^2]/q\to0 E [ p l ( n ) 2 ] / q → 0 uniformly in l l l and n n n .
Proof. For (i), the 2 N 2N 2 N desks of cluster l l l contribute equally, such that ν l = κ l w l / ( 2 N ) \nu_l=\kappa_lw_l/(2N) ν l = κ l w l / ( 2 N ) and ν l / ν ˉ = ϖ l / ∑ k ϖ k 2 \nu_l/\bar{\nu}=\varpi_l/\sum_k\varpi_k^2 ν l / ν ˉ = ϖ l / ∑ k ϖ k 2 (§ 5.2 ). Bounding κ l \kappa_l κ l by inf κ \inf\kappa inf κ and sup κ \sup\kappa sup κ in ϖ l = κ l τ l − α / ∑ j κ j τ j − α \varpi_l=\kappa_l\tau_l^{-\alpha}/\sum_j\kappa_j\tau_j^{-\alpha} ϖ l = κ l τ l − α / ∑ j κ j τ j − α and summing the geometric series in τ k = τ 1 Ω k − 1 \tau_k=\tau_1\Omega^{k-1} τ k = τ 1 Ω k − 1 , we obtain
ν l ν ˉ ≤ ( sup κ inf κ ) 3 τ l − α ∑ j τ j − α ∑ k τ k − 2 α ≤ ( sup κ inf κ ) 3 1 − Ω − 2 α ( 1 − Ω − α ) ( 1 − Ω − 2 α K ) = ( sup κ inf κ ) 3 1 + Ω − α 1 − Ω − 2 α K . \frac{\nu_l}{\bar{\nu}}\le\left(\frac{\sup\kappa}{\inf\kappa}\right)^3\frac{\tau_l^{-\alpha}\sum_j\tau_j^{-\alpha}}{\sum_k\tau_k^{-2\alpha}}\le\left(\frac{\sup\kappa}{\inf\kappa}\right)^3\frac{1-\Omega^{-2\alpha}}{(1-\Omega^{-\alpha})(1-\Omega^{-2\alpha K})}=\left(\frac{\sup\kappa}{\inf\kappa}\right)^3\frac{1+\Omega^{-\alpha}}{1-\Omega^{-2\alpha K}}. ν ˉ ν l ≤ ( inf κ sup κ ) 3 ∑ k τ k − 2 α τ l − α ∑ j τ j − α ≤ ( inf κ sup κ ) 3 ( 1 − Ω − α ) ( 1 − Ω − 2 α K ) 1 − Ω − 2 α = ( inf κ sup κ ) 3 1 − Ω − 2 α K 1 + Ω − α .
As Ω K − 1 = τ K / τ 1 → ∞ \Omega^{K-1}=\tau_K/\tau_1\to\infty Ω K − 1 = τ K / τ 1 → ∞ by (R1), the last fraction is at most 4 4 4 for large T T T .
For (ii), we insert (i) into the bound on E [ I ~ m 2 ] \mathbb{E}[\tilde{I}_m^2] E [ I ~ m 2 ] and this bound into the bound on Var ( E [ I n ∣ F n ] ) \operatorname{Var}(\mathbb{E}[I_n\mid\mathcal{F}_n]) Var ( E [ I n ∣ F n ]) . With I ˉ = p 0 κ ˉ / ( 1 − ∥ ϕ ∥ ) \bar{I}=p_0\bar{\kappa}/(1-\lVert\phi\rVert) I ˉ = p 0 κ ˉ / ( 1 − ∥ ϕ ∥) , we obtain
Var ( E [ I n ∣ F n ] ) I ˉ 2 ≤ C ν ν ˉ I ˉ C R γ ( 1 − ∥ ϕ ∥ ) 2 T = C ν C R γ ν ˉ / κ ˉ T ( 1 − ∥ ϕ ∥ ) p 0 → C ν C R C 2 , \frac{\operatorname{Var}(\mathbb{E}[I_n\mid\mathcal{F}_n])}{\bar{I}^2}\le C_{\nu}\frac{\bar{\nu}}{\bar{I}}C_R\frac{\gamma}{(1-\lVert\phi\rVert)^2T}=C_{\nu}C_R\frac{\gamma\bar{\nu}/\bar{\kappa}}{T(1-\lVert\phi\rVert)p_0}\to\frac{C_{\nu}C_R}{C_2}, I ˉ 2 Var ( E [ I n ∣ F n ]) ≤ C ν I ˉ ν ˉ C R ( 1 − ∥ ϕ ∥ ) 2 T γ = C ν C R T ( 1 − ∥ ϕ ∥) p 0 γ ν ˉ / κ ˉ → C 2 C ν C R ,
where the limit follows from the second condition in (R2) and ∥ ϕ ∥ → 1 \lVert\phi\rVert\to1 ∥ ϕ ∥ → 1 . Adding E [ E [ I n ∣ F n ] 2 ] = E [ I n ] 2 ≤ I ˉ 2 \mathbb{E}[\mathbb{E}[I_n\mid\mathcal{F}_n]^2]=\mathbb{E}[I_n]^2\le\bar{I}^2 E [ E [ I n ∣ F n ] 2 ] = E [ I n ] 2 ≤ I ˉ 2 , we obtain (ii) for large T T T .
For (iii), we split U l U_l U l with I m = E [ I m ∣ F m ] + I ~ m I_m=\mathbb{E}[I_m\mid\mathcal{F}_m]+\tilde{I}_m I m = E [ I m ∣ F m ] + I ~ m . Since the weights ω l \omega_l ω l sum to at most one and the deviations I ~ m \tilde{I}_m I ~ m are uncorrelated, we obtain from Minkowski’s inequality
E [ p l ( n ) 2 ] 1 / 2 ≤ p 0 + β ∑ m = 0 n − 1 ω l ( n − m ) E [ E [ I m ∣ F m ] 2 ] 1 / 2 + β ( ∑ m = 0 n − 1 ω l ( n − m ) 2 E [ I ~ m 2 ] ) 1 / 2 ≤ q + β C λ I ˉ + β ( γ τ l C ν ν ˉ I ˉ ) 1 / 2 , \begin{aligned}
\mathbb{E}[p_l(n)^2]^{1/2}\le p_0+\beta\sum_{m=0}^{n-1}\omega_l(n-m)\mathbb{E}\left[\mathbb{E}[I_m\mid\mathcal{F}_m]^2\right]^{1/2}+\beta\left(\sum_{m=0}^{n-1}\omega_l(n-m)^2\mathbb{E}[\tilde{I}_m^2]\right)^{1/2} \\
&\le q+\beta\sqrt{C_{\lambda}}\bar{I}+\beta\left(\frac{\gamma}{\tau_l}C_{\nu}\bar{\nu}\bar{I}\right)^{1/2},
\end{aligned} E [ p l ( n ) 2 ] 1/2 ≤ p 0 + β m = 0 ∑ n − 1 ω l ( n − m ) E [ E [ I m ∣ F m ] 2 ] 1/2 + β ( m = 0 ∑ n − 1 ω l ( n − m ) 2 E [ I ~ m 2 ] ) 1/2 ≤ q + β C λ I ˉ + β ( τ l γ C ν ν ˉ I ˉ ) 1/2 ,
where we used p 0 ≤ q p_0\le q p 0 ≤ q , (ii), ω l ≤ γ / τ l \omega_l\le\gamma/\tau_l ω l ≤ γ / τ l (Lemma 4.4 ) and the bound on E [ I ~ m 2 ] \mathbb{E}[\tilde{I}_m^2] E [ I ~ m 2 ] above. By Proposition 4.7 , β I ˉ = ∥ ϕ ∥ q ≤ q \beta\bar{I}=\lVert\phi\rVert q\le q β I ˉ = ∥ ϕ ∥ q ≤ q and β 2 ν ˉ I ˉ = ∥ ϕ ∥ 2 q ν ˉ / κ ˉ \beta^2\bar{\nu}\bar{I}=\lVert\phi\rVert^2q\bar{\nu}/\bar{\kappa} β 2 ν ˉ I ˉ = ∥ ϕ ∥ 2 q ν ˉ / κ ˉ . Squaring with ( a + b + c ) 2 ≤ 3 ( a 2 + b 2 + c 2 ) (a+b+c)^2\le3(a^2+b^2+c^2) ( a + b + c ) 2 ≤ 3 ( a 2 + b 2 + c 2 ) and using τ l ≥ τ 1 \tau_l\ge\tau_1 τ l ≥ τ 1 , we obtain (iii). □ \square □
Splitting I n = E [ I n ∣ F n ] + I ~ n I_n=\mathbb{E}[I_n\mid\mathcal{F}_n]+\tilde{I}_n I n = E [ I n ∣ F n ] + I ~ n in X T \mathcal{X}^T X T into the two quantities of Lemma 5.7 , we obtain
X t T = γ T ∑ n ≤ T t / γ E [ I n ∣ F n ] I ˉ + ε T Z t T , Z t T = ( γ T I ˉ ν ˉ ) 1 / 2 ∑ n ≤ T t / γ I ~ n , ε T = ( γ ν ˉ T I ˉ ) 1 / 2 , \mathcal{X}_t^T=\frac{\gamma}{T}\sum_{n\le Tt/\gamma}\frac{\mathbb{E}[I_n\mid\mathcal{F}_n]}{\bar{I}}+\varepsilon_T\mathcal{Z}_t^T,\qquad\mathcal{Z}_t^T=\left(\frac{\gamma}{T\bar{I}\bar{\nu}}\right)^{1/2}\sum_{n\le Tt/\gamma}\tilde{I}_n,\qquad\varepsilon_T=\left(\frac{\gamma\bar{\nu}}{T\bar{I}}\right)^{1/2}, X t T = T γ n ≤ T t / γ ∑ I ˉ E [ I n ∣ F n ] + ε T Z t T , Z t T = ( T I ˉ ν ˉ γ ) 1/2 n ≤ T t / γ ∑ I ~ n , ε T = ( T I ˉ γ ν ˉ ) 1/2 ,
where Z T \mathcal{Z}^T Z T is the rescaled deviation of § 5.1 under the correspondence of § 5.2 . By the second condition in (R2),
ε T 2 = ( 1 − ∥ ϕ ∥ ) γ ν ˉ / κ ˉ T p 0 = ( 1 − ∥ ϕ ∥ ) 2 ∥ ϕ ∥ C 2 ( 1 + o ( 1 ) ) → 0. \varepsilon_T^2=\frac{(1-\lVert\phi\rVert)\gamma\bar{\nu}/\bar{\kappa}}{Tp_0}=\frac{(1-\lVert\phi\rVert)^2}{\lVert\phi\rVert C_2}(1+o(1))\to0. ε T 2 = T p 0 ( 1 − ∥ ϕ ∥) γ ν ˉ / κ ˉ = ∥ ϕ ∥ C 2 ( 1 − ∥ ϕ ∥ ) 2 ( 1 + o ( 1 )) → 0.
Lemma 5.8 (Increments of the size of the updates). Under the assumptions of Theorem 5.3 , for every sufficiently large T \mathcal{T} T and all 0 ≤ u ≤ t ≤ 1 0 \le u \le t \le1 0 ≤ u ≤ t ≤ 1 ,
E [ ( X t T − X u T ) 2 ] ≤ 2 C λ ( t − u + γ T ) 2 + 2 C ν ε T 2 ( t − u + γ T ) . \mathbb{E}\left[\left(\mathcal{X}^{\mathcal{T}}_t-\mathcal{X}^{\mathcal{T}}_u\right)^2\right]\le2C_\lambda\left(t-u+\frac{\gamma}{\mathcal{T}}\right)^2+2C_\nu\varepsilon_{\mathcal{T}}^2\left(t-u+\frac{\gamma}{\mathcal{T}}\right). E [ ( X t T − X u T ) 2 ] ≤ 2 C λ ( t − u + T γ ) 2 + 2 C ν ε T 2 ( t − u + T γ ) . Moreover, for every δ ∈ [ γ / T , 1 ] \delta\in[\gamma/\mathcal{T},1] δ ∈ [ γ / T , 1 ] and η > 0 \eta>0 η > 0 ,
P ( sup 0 ≤ u ≤ t ≤ 1 , t − u ≤ δ ( X t T − X u T ) > η ) ≤ 36 C λ δ + 12 C ν ε T 2 η 2 . \mathbb{P}\left(\sup_{\substack{0\le u\le t\le1,\ t-u\le\delta}}\left(\mathcal{X}^{\mathcal{T}}_t-\mathcal{X}^{\mathcal{T}}_u\right)>\eta\right)\le\frac{36C_\lambda\delta+12C_\nu\varepsilon_{\mathcal{T}}^2}{\eta^2}. P ( 0 ≤ u ≤ t ≤ 1 , t − u ≤ δ sup ( X t T − X u T ) > η ) ≤ η 2 36 C λ δ + 12 C ν ε T 2 .
As the right side vanishes for T → ∞ \mathcal{T}\to\infty T → ∞ and then δ → 0 \delta\to0 δ → 0 , the rescaled size of the updates has no jumps in the limit.
Proof. The interval ( u , t ] (u,t] ( u , t ] holds at most T ( t − u ) / γ + 1 \mathcal{T}(t-u)/\gamma+1 T ( t − u ) / γ + 1 steps n n n with T u / γ < n ≤ T t / γ \mathcal{T}u/\gamma<n\le\mathcal{T}t/\gamma T u / γ < n ≤ T t / γ . For the expected size of the updates, we obtain from Minkowski’s inequality and Lemma 5.7 (ii)
∥ γ T ∑ T u / γ < n ≤ T t / γ E [ I n ∣ F n ] I ˉ ∥ L 2 ≤ γ T ( T ( t − u ) γ + 1 ) C λ = C λ ( t − u + γ T ) . \left\|\frac{\gamma}{\mathcal{T}}\sum_{\mathcal{T}u/\gamma<n\le\mathcal{T}t/\gamma}\frac{\mathbb{E}[I_n\mid\mathcal{F}_n]}{\bar I}\right\|_{L^2}\le\frac{\gamma}{\mathcal{T}}\left(\frac{\mathcal{T}(t-u)}{\gamma}+1\right)\sqrt{C_\lambda}=\sqrt{C_\lambda}\left(t-u+\frac{\gamma}{\mathcal{T}}\right). T γ T u / γ < n ≤ T t / γ ∑ I ˉ E [ I n ∣ F n ] L 2 ≤ T γ ( γ T ( t − u ) + 1 ) C λ = C λ ( t − u + T γ ) .
As the deviations are uncorrelated, we obtain with E [ I ~ n 2 ] ≤ C ν ν ˉ I ˉ \mathbb{E}[\widetilde I_n^2]\le C_\nu\bar\nu\bar I E [ I n 2 ] ≤ C ν ν ˉ I ˉ from Lemma 5.7 (i)
ε T 2 E [ ( Z t T − Z u T ) 2 ] = γ 2 T 2 I ˉ 2 ∑ T u / γ < n ≤ T t / γ E [ I ~ n 2 ] ≤ C ν ε T 2 ( t − u + γ T ) . \varepsilon_{\mathcal{T}}^2\mathbb{E}\left[\left(\mathcal{Z}^{\mathcal{T}}_t-\mathcal{Z}^{\mathcal{T}}_u\right)^2\right]=\frac{\gamma^2}{\mathcal{T}^2\bar I^2}\sum_{\mathcal{T}u/\gamma<n\le\mathcal{T}t/\gamma}\mathbb{E}[\widetilde I_n^2]\le C_\nu\varepsilon_{\mathcal{T}}^2\left(t-u+\frac{\gamma}{\mathcal{T}}\right). ε T 2 E [ ( Z t T − Z u T ) 2 ] = T 2 I ˉ 2 γ 2 T u / γ < n ≤ T t / γ ∑ E [ I n 2 ] ≤ C ν ε T 2 ( t − u + T γ ) .
With ( a + b ) 2 ≤ 2 a 2 + 2 b 2 (a+b)^2\le2a^2+2b^2 ( a + b ) 2 ≤ 2 a 2 + 2 b 2 , we obtain the first claim. For the second claim, let t k = min { k δ , 1 } t_k=\min\{k\delta,1\} t k = min { k δ , 1 } for k = 0 , … , ⌈ 1 / δ ⌉ k=0,\ldots,\lceil1/\delta\rceil k = 0 , … , ⌈ 1/ δ ⌉ . As X T \mathcal{X}^{\mathcal{T}} X T does not decrease, every pair u ≤ t u\le t u ≤ t with t − u ≤ δ t-u\le\delta t − u ≤ δ lies in some [ t k , t k + 2 ] [t_k,t_{k+2}] [ t k , t k + 2 ] , such that
sup t − u ≤ δ ( X t T − X u T ) ≤ max k ( X t k + 2 T − X t k T ) . \sup_{t-u\le\delta}\left(\mathcal{X}^{\mathcal{T}}_t-\mathcal{X}^{\mathcal{T}}_u\right)\le\max_k\left(\mathcal{X}^{\mathcal{T}}_{t_{k+2}}-\mathcal{X}^{\mathcal{T}}_{t_k}\right). t − u ≤ δ sup ( X t T − X u T ) ≤ k max ( X t k + 2 T − X t k T ) .
Summing Chebyshev’s inequality over the at most 2 / δ 2/\delta 2/ δ values of k k k and inserting the first claim with t k + 2 − t k + γ / T ≤ 3 δ t_{k+2}-t_k+\gamma/\mathcal{T}\le3\delta t k + 2 − t k + γ / T ≤ 3 δ , we obtain
P ( max k ( X t k + 2 T − X t k T ) > η ) ≤ 2 δ η 2 ( 18 C λ δ 2 + 6 C ν ε T 2 δ ) = 36 C λ δ + 12 C ν ε T 2 η 2 . \mathbb{P}\left(\max_k\left(\mathcal{X}^{\mathcal{T}}_{t_{k+2}}-\mathcal{X}^{\mathcal{T}}_{t_k}\right)>\eta\right)\le\frac{2}{\delta\eta^2}\left(18C_\lambda\delta^2+6C_\nu\varepsilon_{\mathcal{T}}^2\delta\right)=\frac{36C_\lambda\delta+12C_\nu\varepsilon_{\mathcal{T}}^2}{\eta^2}. P ( k max ( X t k + 2 T − X t k T ) > η ) ≤ δ η 2 2 ( 18 C λ δ 2 + 6 C ν ε T 2 δ ) = η 2 36 C λ δ + 12 C ν ε T 2 .
For the rescaled deviation Z T \mathcal{Z}^{\mathcal{T}} Z T of our market, the process ⟨ Z T ⟩ \langle\mathcal{Z}^{\mathcal{T}}\rangle ⟨ Z T ⟩ of § 5.1 is the sum of the conditional variances of its increments,
⟨ Z T ⟩ t = γ T I ˉ ν ˉ ∑ n ≤ T t / γ Var [ I n ∣ F n ] , t ∈ [ 0 , 1 ] . \langle\mathcal{Z}^{\mathcal{T}}\rangle_t=\frac{\gamma}{\mathcal{T}\bar I\bar\nu}\sum_{n\le\mathcal{T}t/\gamma}\operatorname{Var}[I_n\mid\mathcal{F}_n],\qquad t\in[0,1]. ⟨ Z T ⟩ t = T I ˉ ν ˉ γ n ≤ T t / γ ∑ Var [ I n ∣ F n ] , t ∈ [ 0 , 1 ] .
Let n t = ⌊ T t / γ ⌋ n_t=\lfloor\mathcal{T}t/\gamma\rfloor n t = ⌊ T t / γ ⌋ . Dividing Lemma 5.5 by T I ˉ ν ˉ / γ \mathcal{T}\bar I\bar\nu/\gamma T I ˉ ν ˉ / γ and inserting ( γ / T ) ∑ n ≤ n t E [ I n ∣ F n ] / I ˉ = X t T − ε T Z t T (\gamma/\mathcal{T})\sum_{n\le n_t}\mathbb{E}[I_n\mid\mathcal{F}_n]/\bar I=\mathcal{X}^{\mathcal{T}}_t-\varepsilon_{\mathcal{T}}\mathcal{Z}^{\mathcal{T}}_t ( γ / T ) ∑ n ≤ n t E [ I n ∣ F n ] / I ˉ = X t T − ε T Z t T , we obtain
⟨ Z T ⟩ t = X t T − ε T Z t T − D t T , \langle\mathcal{Z}^{\mathcal{T}}\rangle_t=\mathcal{X}^{\mathcal{T}}_t-\varepsilon_{\mathcal{T}}\mathcal{Z}^{\mathcal{T}}_t-D^{\mathcal{T}}_t, ⟨ Z T ⟩ t = X t T − ε T Z t T − D t T ,
D t T = γ T I ˉ ν ˉ ( ϱ ∑ l = 1 K ( ν l − ν ˉ ) κ l w l ( T l − 1 ) U l ( n t ) + ∑ l = 1 K ν l κ l w l ∑ r = 0 n t p l ( r ) 2 ) . D^{\mathcal{T}}_t=\frac{\gamma}{\mathcal{T}\bar I\bar\nu}\left(\varrho\sum_{l=1}^{K}(\nu_l-\bar\nu)\kappa_lw_l(\mathcal{T}_l-1)U_l(n_t)+\sum_{l=1}^{K}\nu_l\kappa_lw_l\sum_{r=0}^{n_t}p_l(r)^2\right). D t T = T I ˉ ν ˉ γ ( ϱ l = 1 ∑ K ( ν l − ν ˉ ) κ l w l ( T l − 1 ) U l ( n t ) + l = 1 ∑ K ν l κ l w l r = 0 ∑ n t p l ( r ) 2 ) .
Lemma 5.9 (Conditional variance in the limit). Under the assumptions of Theorem 5.3 , as T → ∞ \mathcal{T}\to\infty T → ∞ ,
sup t ≤ 1 ∣ D t T ∣ → 0 and sup t ≤ 1 ∣ ⟨ Z T ⟩ t − X t T ∣ → 0 in probability . \sup_{t\le1}|D^{\mathcal{T}}_t|\to0\qquad\text{and}\qquad\sup_{t\le1}\left|\langle\mathcal{Z}^{\mathcal{T}}\rangle_t-\mathcal{X}^{\mathcal{T}}_t\right|\to0\quad\text{in probability}. t ≤ 1 sup ∣ D t T ∣ → 0 and t ≤ 1 sup ⟨ Z T ⟩ t − X t T → 0 in probability .
In units of ν ˉ \bar{\nu} ν ˉ , the conditional variance of the size of the updates thus equals its conditional mean in the limit, as for a number of actions.
Proof. We first bound the part of D T D^{\mathcal{T}} D T with the response after the step n t n_t n t . Let Δ X m = γ I m / ( T I ˉ ) \Delta X_m=\gamma I_m/(T\bar{I}) Δ X m = γ I m / ( T I ˉ ) denote the increments of X T X^{\mathcal{T}} X T . Using ∣ ν l − ν ˉ ∣ ≤ C ν ν ˉ |\nu_l-\bar{\nu}|\le C_\nu\bar{\nu} ∣ ν l − ν ˉ ∣ ≤ C ν ν ˉ (Lemma 5.7 (i)), β κ l w l = ∥ ϕ ∥ ϖ l \beta\kappa_lw_l=\|\phi\|\varpi_l β κ l w l = ∥ ϕ ∥ ϖ l (Proposition 4.7 ) and the proof of Lemma 5.4 , we obtain
γ T I ˉ β ∑ l = 1 K ∣ ν l − ν ˉ ∣ κ l w l ( T l − 1 ) U l ( n t ) ≤ C ν ∑ l = 1 K ϖ l S l ( t ) , S l ( t ) = ∑ m < n t e − ( n t − m ) γ / τ l Δ X m . \frac{\gamma}{\mathcal{T}\bar{I}}\beta\sum_{l=1}^{K}|\nu_l-\bar{\nu}|\kappa_lw_l(\mathcal{T}_l-1)U_l(n_t)\le C_\nu\sum_{l=1}^{K}\varpi_lS_l(t),\qquad S_l(t)=\sum_{m<n_t}e^{-(n_t-m)\gamma/\tau_l}\Delta X_m. T I ˉ γ β l = 1 ∑ K ∣ ν l − ν ˉ ∣ κ l w l ( T l − 1 ) U l ( n t ) ≤ C ν l = 1 ∑ K ϖ l S l ( t ) , S l ( t ) = m < n t ∑ e − ( n t − m ) γ / τ l Δ X m .
For δ ∈ [ γ / T , 1 ] \delta\in[\gamma/T,1] δ ∈ [ γ / T , 1 ] , the increments within the time δ \delta δ before t t t sum to at most the largest increment w T ( 2 δ ) w_T(2\delta) w T ( 2 δ ) of X T X^{\mathcal{T}} X T over intervals of length 2 δ 2\delta 2 δ , while the earlier increments carry the weight at most e − δ T / τ l e^{-\delta T/\tau_l} e − δ T / τ l . Bounding the weights of the long horizons by the geometric series in τ l − α \tau_l^{-\alpha} τ l − α as in the proof of Lemma 5.7 (i), we thus obtain
∑ l = 1 K ϖ l S l ( t ) ≤ w T ( 2 δ ) + X 1 T ( e − 1 / δ + ∑ l : τ l > δ 2 T ϖ l ) ≤ w T ( 2 δ ) + X 1 T ( e − 1 / δ + 2 sup κ inf κ ( τ 1 δ 2 T ) α ) . \sum_{l=1}^{K}\varpi_lS_l(t)\le w_T(2\delta)+\mathcal{X}_1^{\mathcal{T}}\left(e^{-1/\delta}+\sum_{l:\tau_l>\delta^2T}\varpi_l\right)\le w_T(2\delta)+\mathcal{X}_1^{\mathcal{T}}\left(e^{-1/\delta}+2\frac{\sup\kappa}{\inf\kappa}\left(\frac{\tau_1}{\delta^2T}\right)^\alpha\right). l = 1 ∑ K ϖ l S l ( t ) ≤ w T ( 2 δ ) + X 1 T e − 1/ δ + l : τ l > δ 2 T ∑ ϖ l ≤ w T ( 2 δ ) + X 1 T ( e − 1/ δ + 2 inf κ sup κ ( δ 2 T τ 1 ) α ) .
The right side does not depend on t t t . By Lemma 5.8 , P ( w T ( 2 δ ) > η ) ≤ ( 72 C λ δ + 12 C ν ε T 2 ) / η 2 \mathbb{P}(w_T(2\delta)>\eta)\le(72C_\lambda\delta+12C_\nu\varepsilon_T^2)/\eta^2 P ( w T ( 2 δ ) > η ) ≤ ( 72 C λ δ + 12 C ν ε T 2 ) / η 2 for every η > 0 \eta>0 η > 0 , while E [ X 1 T ] ≤ 1 + γ / T \mathbb{E}[\mathcal{X}_1^{\mathcal{T}}]\le1+\gamma/T E [ X 1 T ] ≤ 1 + γ / T by Proposition 4.8 . Letting first T → ∞ T\to\infty T → ∞ and then δ → 0 \delta\to0 δ → 0 , we obtain that this part of sup t ∣ D t T ∣ \sup_{t}|D_t^{\mathcal{T}}| sup t ∣ D t T ∣ tends to zero in probability.
The part of D T D^{\mathcal{T}} D T with the squared probabilities of acting does not decrease in t t t . With ν l ≤ C ν ν ˉ \nu_l\le C_\nu\bar{\nu} ν l ≤ C ν ν ˉ , κ ˉ / I ˉ = 1 / q \bar{\kappa}/\bar{I}=1/q κ ˉ / I ˉ = 1/ q and Lemma 5.7 (iii), we obtain
E [ γ T I ˉ ν ˉ ∑ l = 1 K ν l κ l w l ∑ r = 0 n 1 p l ( r ) 2 ] ≤ C ν ( 1 + γ T ) max l , r E [ p l ( r ) 2 ] q ⟶ 0 , \mathbb{E}\left[\frac{\gamma}{T\bar{I}\bar{\nu}}\sum_{l=1}^{K}\nu_l\kappa_lw_l\sum_{r=0}^{n_1}p_l(r)^2\right]\le C_\nu\left(1+\frac{\gamma}{T}\right)\max_{l,r}\frac{\mathbb{E}[p_l(r)^2]}{q}\longrightarrow0, E [ T I ˉ ν ˉ γ l = 1 ∑ K ν l κ l w l r = 0 ∑ n 1 p l ( r ) 2 ] ≤ C ν ( 1 + T γ ) l , r max q E [ p l ( r ) 2 ] ⟶ 0 ,
such that this part tends to zero in probability by Markov’s inequality. Lastly, we obtain from Doob’s inequality and the proof of Lemma 5.8 that E [ sup t ≤ 1 ∣ Z t T ∣ 2 ] ≤ 4 E [ ( Z 1 T ) 2 ] ≤ 4 C ν ( 1 + γ / T ) \mathbb{E}[\sup_{t\le1}|Z_t^{\mathcal{T}}|^2]\le4\mathbb{E}[(Z_1^{\mathcal{T}})^2]\le4C_\nu(1+\gamma/T) E [ sup t ≤ 1 ∣ Z t T ∣ 2 ] ≤ 4 E [( Z 1 T ) 2 ] ≤ 4 C ν ( 1 + γ / T ) . As ε T → 0 \varepsilon_T\to0 ε T → 0 , the term ε T Z T \varepsilon_TZ^{\mathcal{T}} ε T Z T thus tends to zero uniformly in probability, and the second claim follows from the first.
The second identity of § 5.1 , [ Z T , Z T ] = X T [Z^{\mathcal{T}},Z^{\mathcal{T}}]=X^{\mathcal{T}} [ Z T , Z T ] = X T , holds because the number of actions jumps by one. In our market, the jump of Z T Z^{\mathcal{T}} Z T in step n n n is ( γ / ( T I ˉ ν ˉ ) ) 1 / 2 I ~ n (\gamma/(T\bar{I}\bar{\nu}))^{1/2}\tilde{I}_n ( γ / ( T I ˉ ν ˉ ) ) 1/2 I ~ n . The sum of the squared jumps of Z T Z^{\mathcal{T}} Z T is therefore
[ Z T , Z T ] t = γ T I ˉ ν ˉ ∑ n ≤ T t / γ I ~ n 2 . [Z^{\mathcal{T}},Z^{\mathcal{T}}]_t=\frac{\gamma}{T\bar{I}\bar{\nu}}\sum_{n\le Tt/\gamma}\tilde{I}_n^2. [ Z T , Z T ] t = T I ˉ ν ˉ γ n ≤ T t / γ ∑ I ~ n 2 .
Subtracting the sum of the conditional variances ⟨ Z T ⟩ \langle Z^{\mathcal{T}}\rangle ⟨ Z T ⟩ , we obtain
[ Z T , Z T ] t − ⟨ Z T ⟩ t = γ T I ˉ ν ˉ ∑ n ≤ T t / γ ( I ~ n 2 − Var [ I n ∣ F n ] ) . [Z^{\mathcal{T}},Z^{\mathcal{T}}]_t-\langle Z^{\mathcal{T}}\rangle_t=\frac{\gamma}{T\bar{I}\bar{\nu}}\sum_{n\le Tt/\gamma}\left(\tilde{I}_n^2-\operatorname{Var}[I_n\mid\mathcal{F}_n]\right). [ Z T , Z T ] t − ⟨ Z T ⟩ t = T I ˉ ν ˉ γ n ≤ T t / γ ∑ ( I ~ n 2 − Var [ I n ∣ F n ] ) .
Lemma 5.10 (Squared jumps in the limit). Under the assumptions of Theorem 5.3 , as T → ∞ T\to\infty T → ∞ ,
E [ sup t ≤ 1 ( [ Z T , Z T ] t − ⟨ Z T ⟩ t ) 2 ] ⟶ 0 , max n ≤ T / γ ( γ T I ˉ ν ˉ ) 1 / 2 ∣ I ~ n ∣ ⟶ 0 , max n ≤ T / γ γ I n T I ˉ ⟶ 0 , \mathbb{E}\left[\sup_{t\le1}\left([Z^{\mathcal{T}},Z^{\mathcal{T}}]_t-\langle Z^{\mathcal{T}}\rangle_t\right)^2\right]\longrightarrow0,\qquad\max_{n\le T/\gamma}\left(\frac{\gamma}{T\bar{I}\bar{\nu}}\right)^{1/2}|\tilde{I}_n|\longrightarrow0,\qquad\max_{n\le T/\gamma}\frac{\gamma I_n}{T\bar{I}}\longrightarrow0, E [ t ≤ 1 sup ( [ Z T , Z T ] t − ⟨ Z T ⟩ t ) 2 ] ⟶ 0 , n ≤ T / γ max ( T I ˉ ν ˉ γ ) 1/2 ∣ I ~ n ∣ ⟶ 0 , n ≤ T / γ max T I ˉ γ I n ⟶ 0 , where the last two limits hold in probability. In particular, sup t ≤ 1 ∣ [ Z T , Z T ] t − X t T ∣ ⟶ 0 \sup_{t\le1}|[Z^{\mathcal{T}},Z^{\mathcal{T}}]_t-X_t^{\mathcal{T}}|\longrightarrow0 sup t ≤ 1 ∣ [ Z T , Z T ] t − X t T ∣ ⟶ 0 in probability.
In the limit, the size of the updates thus satisfies both identities of § 5.1 , while the jumps of Z T \mathcal{Z}^{\mathcal{T}} Z T and X T \mathcal{X}^{\mathcal{T}} X T vanish.
Proof. Given F n \mathcal{F}_n F n , the deviation I ~ n \widetilde{I}_n I n is the sum of the independent contributions ν l ( ξ i ( n ) − p l ( n ) ) \nu_l(\xi_i(n)-p_l(n)) ν l ( ξ i ( n ) − p l ( n )) of the desks, which have mean zero and satisfy ∣ ν l ( ξ i ( n ) − p l ( n ) ) ∣ ≤ C ν ν ˉ |\nu_l(\xi_i(n)-p_l(n))| \le C_\nu\bar{\nu} ∣ ν l ( ξ i ( n ) − p l ( n )) ∣ ≤ C ν ν ˉ by Lemma 5.7 (i). Expanding the fourth power and using Var [ I n ∣ F n ] ≤ C ν ν ˉ E [ I n ∣ F n ] \operatorname{Var}[I_n \mid\mathcal{F}_n] \le C_\nu\bar{\nu}\mathbb{E}[I_n \mid\mathcal{F}_n] Var [ I n ∣ F n ] ≤ C ν ν ˉ E [ I n ∣ F n ] , we obtain
E [ I ~ n 4 ∣ F n ] ≤ C ν 2 ν ˉ 2 Var [ I n ∣ F n ] + 3 Var [ I n ∣ F n ] 2 ≤ C ν 3 ν ˉ 3 E [ I n ∣ F n ] + 3 C ν 2 ν ˉ 2 E [ I n ∣ F n ] 2 . \begin{aligned}
\mathbb{E}[\widetilde{I}_n^4 \mid\mathcal{F}_n] \le C_\nu^2\bar{\nu}^2\operatorname{Var}[I_n \mid\mathcal{F}_n] + 3\operatorname{Var}[I_n \mid\mathcal{F}_n]^2 \\
&\le C_\nu^3\bar{\nu}^3\mathbb{E}[I_n \mid\mathcal{F}_n] + 3C_\nu^2\bar{\nu}^2\mathbb{E}[I_n \mid\mathcal{F}_n]^2.
\end{aligned} E [ I n 4 ∣ F n ] ≤ C ν 2 ν ˉ 2 Var [ I n ∣ F n ] + 3 Var [ I n ∣ F n ] 2 ≤ C ν 3 ν ˉ 3 E [ I n ∣ F n ] + 3 C ν 2 ν ˉ 2 E [ I n ∣ F n ] 2 .
Taking expectations with E [ I n ] ≤ I ˉ \mathbb{E}[I_n] \le\bar{I} E [ I n ] ≤ I ˉ and Lemma 5.7 (ii), and summing over the at most T / γ + 1 \mathcal{T}/\gamma+1 T / γ + 1 steps, we obtain with ε T 2 = γ ν ˉ / ( T I ˉ ) \varepsilon_{\mathcal{T}}^2=\gamma\bar{\nu}/(\mathcal{T}\bar{I}) ε T 2 = γ ν ˉ / ( T I ˉ )
( γ T I ˉ ν ˉ ) 2 ∑ n ≤ T / γ E [ I ~ n 4 ] ≤ ( 1 + γ T ) ( C ν 3 ε T 2 + 3 C ν 2 C λ γ T ) → 0. \left(\frac{\gamma}{\mathcal{T}\bar{I}\bar{\nu}}\right)^2\sum_{n\le\mathcal{T}/\gamma}\mathbb{E}[\widetilde{I}_n^4]\le\left(1+\frac{\gamma}{\mathcal{T}}\right)\left(C_\nu^3\varepsilon_{\mathcal{T}}^2+3C_\nu^2C_\lambda\frac{\gamma}{\mathcal{T}}\right)\to0. ( T I ˉ ν ˉ γ ) 2 n ≤ T / γ ∑ E [ I n 4 ] ≤ ( 1 + T γ ) ( C ν 3 ε T 2 + 3 C ν 2 C λ T γ ) → 0.
The differences I ~ n 2 − Var [ I n ∣ F n ] \widetilde{I}_n^2-\operatorname{Var}[I_n\mid\mathcal{F}_n] I n 2 − Var [ I n ∣ F n ] have conditional mean zero and second moment at most E [ I ~ n 4 ] \mathbb{E}[\widetilde{I}_n^4] E [ I n 4 ] . By Doob’s inequality, the expectation in the first limit is thus at most four times this sum of fourth moments. Moreover, we obtain from Markov’s inequality with the fourth power
P ( max n ≤ T / γ ( γ T I ˉ ν ˉ ) 1 / 2 ∣ I ~ n ∣ > η ) ≤ 1 η 4 ( γ T I ˉ ν ˉ ) 2 ∑ n ≤ T / γ E [ I ~ n 4 ] → 0. \mathbb{P}\left(\max_{n\le\mathcal{T}/\gamma}\left(\frac{\gamma}{\mathcal{T}\bar{I}\bar{\nu}}\right)^{1/2}|\widetilde{I}_n|>\eta\right)\le\frac{1}{\eta^4}\left(\frac{\gamma}{\mathcal{T}\bar{I}\bar{\nu}}\right)^2\sum_{n\le\mathcal{T}/\gamma}\mathbb{E}[\widetilde{I}_n^4]\to0. P ( n ≤ T / γ max ( T I ˉ ν ˉ γ ) 1/2 ∣ I n ∣ > η ) ≤ η 4 1 ( T I ˉ ν ˉ γ ) 2 n ≤ T / γ ∑ E [ I n 4 ] → 0.
For the jumps of X T \mathcal{X}^{\mathcal{T}} X T , we write γ I n / ( T I ˉ ) = γ E [ I n ∣ F n ] / ( T I ˉ ) + ε T ( γ / ( T I ˉ ν ˉ ) ) 1 / 2 I ~ n \gamma I_n/(\mathcal{T}\bar{I})=\gamma\mathbb{E}[I_n\mid\mathcal{F}_n]/(\mathcal{T}\bar{I})+\varepsilon_{\mathcal{T}}(\gamma/(\mathcal{T}\bar{I}\bar{\nu}))^{1/2}\widetilde{I}_n γ I n / ( T I ˉ ) = γ E [ I n ∣ F n ] / ( T I ˉ ) + ε T ( γ / ( T I ˉ ν ˉ ) ) 1/2 I n . The second term vanishes by the previous limit and ε T → 0 \varepsilon_{\mathcal{T}}\to0 ε T → 0 , while we obtain from Chebyshev’s inequality and Lemma 5.7 (ii)
P ( max n ≤ T / γ γ E [ I n ∣ F n ] T I ˉ > η ) ≤ 1 η 2 ( T γ + 1 ) γ 2 T 2 C λ = C λ η 2 γ T ( 1 + γ T ) → 0. \mathbb{P}\left(\max_{n\le\mathcal{T}/\gamma}\frac{\gamma\mathbb{E}[I_n\mid\mathcal{F}_n]}{\mathcal{T}\bar{I}}>\eta\right)\le\frac{1}{\eta^2}\left(\frac{\mathcal{T}}{\gamma}+1\right)\frac{\gamma^2}{\mathcal{T}^2}C_\lambda=\frac{C_\lambda}{\eta^2}\frac{\gamma}{\mathcal{T}}\left(1+\frac{\gamma}{\mathcal{T}}\right)\to0. P ( n ≤ T / γ max T I ˉ γ E [ I n ∣ F n ] > η ) ≤ η 2 1 ( γ T + 1 ) T 2 γ 2 C λ = η 2 C λ T γ ( 1 + T γ ) → 0.
Lastly, the last claim follows from the first limit and Lemma 5.9 . □
For our market, the normalized response ρ T \rho^{\mathcal{T}} ρ T of § 5.1 becomes the total of R R R up to the time T x \mathcal{T}x T x in units of ∑ s R ( s ) = ∥ φ ∥ / ( 1 − ∥ φ ∥ ) \sum_sR(s)=\|\varphi\|/(1-\|\varphi\|) ∑ s R ( s ) = ∥ φ ∥/ ( 1 − ∥ φ ∥ ) (Proposition 4.8 ),
F T ( x ) = 1 − ∥ φ ∥ ∥ φ ∥ ∑ s ≤ T x / γ R ( s ) , x ≥ 0. F^{\mathcal{T}}(x)=\frac{1-\|\varphi\|}{\|\varphi\|}\sum_{s\le\mathcal{T}x/\gamma}R(s),\qquad x\ge0. F T ( x ) = ∥ φ ∥ 1 − ∥ φ ∥ s ≤ T x / γ ∑ R ( s ) , x ≥ 0.
Lemma 5.11 (Response in the limit). Under the assumptions of Theorem 5.3 , let F ( x ) = ∫ 0 x G ( y ) d y F(x)=\int_0^x\mathcal{G}(y)\,dy F ( x ) = ∫ 0 x G ( y ) d y . Then, as T → ∞ \mathcal{T}\to\infty T → ∞ ,
sup x ∈ [ 0 , 1 ] ∣ F T ( x ) − F ( x ) ∣ → 0. \sup_{x\in[0,1]}|F^{\mathcal{T}}(x)-F(x)|\to0. x ∈ [ 0 , 1 ] sup ∣ F T ( x ) − F ( x ) ∣ → 0.
On the time scale T \mathcal{T} T , the response of our market thus has the same limit G \mathcal{G} G as the response of the Hawkes processes in § 5.1 .
Proof. Let z > 0 z>0 z > 0 , w = e − z γ / T w=e^{-z\gamma/\mathcal{T}} w = e − z γ / T and h ( y ) = y / ( 1 + y ) h(y)=y/(1+y) h ( y ) = y / ( 1 + y ) . By the proof of Lemma 5.6 , R ^ = φ ^ / ( 1 − φ ^ ) \widehat{R}=\widehat{\varphi}/(1-\widehat{\varphi}) R = φ / ( 1 − φ ) with 1 − φ ^ ( w ) / ∥ φ ∥ = ∑ l ϖ l ( 1 − w ) / ( 1 − a l w ) 1-\widehat{\varphi}(w)/\|\varphi\|=\sum_l\varpi_l(1-w)/(1-a_lw) 1 − φ ( w ) /∥ φ ∥ = ∑ l ϖ l ( 1 − w ) / ( 1 − a l w ) . Dividing the numerator and the denominator of each term by ( 1 − a l ) w = γ w / τ l (1-a_l)w=\gamma w/\tau_l ( 1 − a l ) w = γ w / τ l , we obtain
1 − φ ^ ( w ) ∥ φ ∥ = ∑ l = 1 K ϖ l ( 1 − w ) / ( γ w / τ l ) ( 1 − w ) / ( γ w / τ l ) + 1 = ∑ l = 1 K ϖ l h ( z T τ l T ) = : S T ( z T ) , z T = T γ ( e z γ / T − 1 ) . (17) 1-\frac{\widehat{\varphi}(w)}{\|\varphi\|}=\sum_{l=1}^{K}\varpi_l\frac{(1-w)/(\gamma w/\tau_l)}{(1-w)/(\gamma w/\tau_l)+1}=\sum_{l=1}^{K}\varpi_l h\left(\frac{z_{\mathcal{T}}\tau_l}{\mathcal{T}}\right)=:S_{\mathcal{T}}(z_{\mathcal{T}}),\qquad z_{\mathcal{T}}=\frac{\mathcal{T}}{\gamma}\left(e^{z\gamma/\mathcal{T}}-1\right). \tag*{(17)} 1 − ∥ φ ∥ φ ( w ) = l = 1 ∑ K ϖ l ( 1 − w ) / ( γ w / τ l ) + 1 ( 1 − w ) / ( γ w / τ l ) = l = 1 ∑ K ϖ l h ( T z T τ l ) =: S T ( z T ) , z T = γ T ( e z γ / T − 1 ) . (17)
With 1 − ∥ ϕ ∥ = ∥ ϕ ∥ C 0 T − α 1-\lVert\phi\rVert=\lVert\phi\rVert C_0\mathcal{T}^{-\alpha} 1 − ∥ ϕ ∥ = ∥ ϕ ∥ C 0 T − α , the Laplace transform of the measure with the distribution function F T F^{\mathcal{T}} F T thus becomes
∑ s ∈ N ∖ { 0 } e − z s γ / T 1 − ∥ ϕ ∥ ∥ ϕ ∥ R ( s ) = 1 − ∥ ϕ ∥ ∥ ϕ ∥ ϕ ^ ( w ) 1 − ϕ ^ ( w ) = 1 − S T ( z T ) 1 + C 0 − 1 T α S T ( z T ) . \sum_{s\in\mathbb{N}\setminus\{0\}}e^{-zs\gamma/\mathcal{T}}\frac{1-\lVert\phi\rVert}{\lVert\phi\rVert}R(s)=\frac{1-\lVert\phi\rVert}{\lVert\phi\rVert}\frac{\hat{\phi}(w)}{1-\hat{\phi}(w)}=\frac{1-S_{\mathcal{T}}(z\mathcal{T})}{1+C_0^{-1}\mathcal{T}^{\alpha}S_{\mathcal{T}}(z\mathcal{T})}. s ∈ N ∖ { 0 } ∑ e − z s γ / T ∥ ϕ ∥ 1 − ∥ ϕ ∥ R ( s ) = ∥ ϕ ∥ 1 − ∥ ϕ ∥ 1 − ϕ ^ ( w ) ϕ ^ ( w ) = 1 + C 0 − 1 T α S T ( z T ) 1 − S T ( z T ) .
We next show that T α S T ( z ′ ) → C 0 z ′ α \mathcal{T}^{\alpha}S_{\mathcal{T}}(z')\to C_0z'^{\alpha} T α S T ( z ′ ) → C 0 z ′ α for every z ′ > 0 z'>0 z ′ > 0 . By the proof of Proposition 4.11 , ϖ l = κ l τ l − α / ∑ j κ j τ j − α \varpi_l=\kappa_l\tau_l^{-\alpha}/\sum_j\kappa_j\tau_j^{-\alpha} ϖ l = κ l τ l − α / ∑ j κ j τ j − α with κ l \kappa_l κ l proportional to κ ( τ l ) \kappa(\tau_l) κ ( τ l ) . Multiplying the numerator and the denominator by ln Ω \ln\Omega ln Ω and writing x l = τ l / T x_l=\tau_l/\mathcal{T} x l = τ l / T , we obtain
T α S T ( z ′ ) = ln Ω ∑ l = 1 K κ ( T x l ) x l − α h ( z ′ x l ) ln Ω ∑ j = 1 K κ ( τ j ) τ j − α . \mathcal{T}^{\alpha}S_{\mathcal{T}}(z')=\frac{\ln\Omega\sum_{l=1}^{K}\kappa(\mathcal{T}x_l)x_l^{-\alpha}h(z'x_l)}{\ln\Omega\sum_{j=1}^{K}\kappa(\tau_j)\tau_j^{-\alpha}}. T α S T ( z ′ ) = ln Ω ∑ j = 1 K κ ( τ j ) τ j − α ln Ω ∑ l = 1 K κ ( T x l ) x l − α h ( z ′ x l ) .
By the proof of Proposition 4.11 , the denominator converges to ∫ τ 1 ∞ κ ( τ ) τ − 1 − α d τ = κ ˉ ∞ / ( α τ 1 α ) \int_{\tau_1}^{\infty}\kappa(\tau)\tau^{-1-\alpha}\,d\tau=\bar{\kappa}_{\infty}/(\alpha\tau_1^{\alpha}) ∫ τ 1 ∞ κ ( τ ) τ − 1 − α d τ = κ ˉ ∞ / ( α τ 1 α ) . The numerator is a Riemann sum with the step ln Ω \ln\Omega ln Ω in ln x l \ln x_l ln x l , over x l x_l x l from τ 1 / T → 0 \tau_1/\mathcal{T}\to0 τ 1 / T → 0 to τ K / T → ∞ \tau_K/\mathcal{T}\to\infty τ K / T → ∞ by (R1). For 0 < x 0 < X 0<x_0<X 0 < x 0 < X , κ ( T x ) → κ ∞ \kappa(\mathcal{T}x)\to\kappa_{\infty} κ ( T x ) → κ ∞ uniformly on [ x 0 , X ] [x_0,X] [ x 0 , X ] , such that the terms with x l ∈ [ x 0 , X ] x_l\in[x_0,X] x l ∈ [ x 0 , X ] converge to κ ∞ ∫ x 0 X x − α h ( z ′ x ) d x / x \kappa_{\infty}\int_{x_0}^{X}x^{-\alpha}h(z'x)\,dx/x κ ∞ ∫ x 0 X x − α h ( z ′ x ) d x / x . As h ( z ′ x ) ≤ z ′ x h(z'x)\le z'x h ( z ′ x ) ≤ z ′ x and h ≤ 1 h\le1 h ≤ 1 , the remaining terms satisfy
ln Ω ∑ l : x l < x 0 κ ( T x l ) x l − α h ( z ′ x l ) ≤ sup κ z ′ ln Ω 1 − Ω − ( 1 − α ) x 0 1 − α , \ln\Omega\sum_{l:x_l<x_0}\kappa(\mathcal{T}x_l)x_l^{-\alpha}h(z'x_l)\le\sup\kappa\frac{z'\ln\Omega}{1-\Omega^{-(1-\alpha)}}x_0^{1-\alpha}, ln Ω l : x l < x 0 ∑ κ ( T x l ) x l − α h ( z ′ x l ) ≤ sup κ 1 − Ω − ( 1 − α ) z ′ ln Ω x 0 1 − α ,
ln Ω ∑ l : x l > X κ ( T x l ) x l − α h ( z ′ x l ) ≤ sup κ ln Ω 1 − Ω − α X − α , \ln\Omega\sum_{l:x_l>X}\kappa(\mathcal{T}x_l)x_l^{-\alpha}h(z'x_l)\le\sup\kappa\frac{\ln\Omega}{1-\Omega^{-\alpha}}X^{-\alpha}, ln Ω l : x l > X ∑ κ ( T x l ) x l − α h ( z ′ x l ) ≤ sup κ 1 − Ω − α ln Ω X − α ,
where ln Ω / ( 1 − Ω − β ) → 1 / β \ln\Omega/(1-\Omega^{-\beta})\to1/\beta ln Ω/ ( 1 − Ω − β ) → 1/ β as Ω → 1 + \Omega\to1^+ Ω → 1 + . Letting T → ∞ \mathcal{T}\to\infty T → ∞ , then x 0 → 0 x_0\to0 x 0 → 0 and X → ∞ X\to\infty X → ∞ , substituting y = z ′ x y=z'x y = z ′ x and inserting k 0 = α τ 1 α κ ∞ Γ ( 1 + α ) / κ ˉ ∞ k_0=\alpha\tau_1^{\alpha}\kappa_{\infty}\Gamma(1+\alpha)/\bar{\kappa}_{\infty} k 0 = α τ 1 α κ ∞ Γ ( 1 + α ) / κ ˉ ∞ (Proposition 4.11 ) into C 0 = ( k 0 / α ) Γ ( 1 − α ) C_0=(k_0/\alpha)\Gamma(1-\alpha) C 0 = ( k 0 / α ) Γ ( 1 − α ) (§ 5.1 ), we obtain
T α S T ( z ′ ) → α τ 1 α κ ∞ κ ˉ ∞ ∫ 0 ∞ x − α h ( z ′ x ) d x x = α τ 1 α κ ∞ κ ˉ ∞ Γ ( α ) Γ ( 1 − α ) z ′ α = C 0 z ′ α . \mathcal{T}^{\alpha}S_{\mathcal{T}}(z')\to\frac{\alpha\tau_1^{\alpha}\kappa_{\infty}}{\bar{\kappa}_{\infty}}\int_0^{\infty}x^{-\alpha}h(z'x)\frac{dx}{x}=\frac{\alpha\tau_1^{\alpha}\kappa_{\infty}}{\bar{\kappa}_{\infty}}\Gamma(\alpha)\Gamma(1-\alpha)z'^{\alpha}=C_0z'^{\alpha}. T α S T ( z ′ ) → κ ˉ ∞ α τ 1 α κ ∞ ∫ 0 ∞ x − α h ( z ′ x ) x d x = κ ˉ ∞ α τ 1 α κ ∞ Γ ( α ) Γ ( 1 − α ) z ′ α = C 0 z ′ α .
Lastly, S T S_{\mathcal{T}} S T increases in its argument, and z T → z z_{\mathcal{T}}\to z z T → z as γ / T → 0 \gamma/\mathcal{T}\to0 γ / T → 0 . For every η ∈ ( 0 , 1 ) \eta\in(0,1) η ∈ ( 0 , 1 ) , we thus have C 0 ( 1 − η ) α z α ≤ lim inf T α S T ( z T ) ≤ lim sup T α S T ( z T ) ≤ C 0 ( 1 + η ) α z α C_0(1-\eta)^{\alpha}z^{\alpha}\le\liminf\mathcal{T}^{\alpha}S_{\mathcal{T}}(z_{\mathcal{T}})\le\limsup\mathcal{T}^{\alpha}S_{\mathcal{T}}(z_{\mathcal{T}})\le C_0(1+\eta)^{\alpha}z^{\alpha} C 0 ( 1 − η ) α z α ≤ lim inf T α S T ( z T ) ≤ lim sup T α S T ( z T ) ≤ C 0 ( 1 + η ) α z α , while S T ( z T ) → 0 S_{\mathcal{T}}(z_{\mathcal{T}})\to0 S T ( z T ) → 0 . The Laplace transform above therefore converges to 1 / ( 1 + z α ) 1/(1+z^{\alpha}) 1/ ( 1 + z α ) , which is the Laplace transform of G G G (§ 5.1 ). By the continuity theorem for Laplace transforms [24 Section XIII.1] , the measures with the distribution functions F T F^{\mathcal{T}} F T thus converge weakly to the measure with the density G G G . As F F F is continuous, F T F^{\mathcal{T}} F T then converges to F F F uniformly. □
Proof of Theorem 5.3 . We follow the proof of Theorem 5.2 in [41 Sections 4.1–4.4] , where the Hawkes process enters in three places. The bound on the mean of N T N^{\mathcal{T}} N T [41 Lemma 4.1] becomes the bound on the increments of X T X^{\mathcal{T}} X T in Lemma 5.8. The two identities for M T M^{\mathcal{T}} M T (§ 5.1 ) become their limits in Lemmas 5.9 and 5.10. The limit of the normalized response [41 Lemma 4.3] becomes Lemma 5.11. With sums over steps in place of integrals over time, the remaining steps apply unchanged. By the correspondence of § 5.2 , the limit has the constants C 1 = 1 C_1=1 C 1 = 1 and C 2 C_2 C 2 of (R2). □
The mid price. In step n n n , each desk d i d_i d i with ξ i ( n ) = 1 \xi_i(n)=1 ξ i ( n ) = 1 changes Z σ i Z^{\sigma_i} Z σ i by m i Δ Z i ( n ) m_i\Delta Z_i(n) m i Δ Z i ( n ) (§ 2.4 ). By (2) , the next move of the mid price is therefore
θ n + 1 = − c ( Z ~ n + − Z ~ n − ) − c ∑ i σ i ξ i ( n ) m i Δ Z i ( n ) , \theta_{n+1}=-c\left(\widetilde{Z}_n^{+}-\widetilde{Z}_n^{-}\right)-c\sum_i\sigma_i\xi_i(n)m_i\Delta Z_i(n), θ n + 1 = − c ( Z n + − Z n − ) − c i ∑ σ i ξ i ( n ) m i Δ Z i ( n ) ,
where Z ~ n σ \widetilde{Z}_n^{\sigma} Z n σ is Z σ Z^{\sigma} Z σ after the shift of step n n n (§ 2.6 ).
Corollary 5.12 (Volatility of the mid price). Under the assumptions of Theorem 5.3 , let
P t T = ( γ T I ‾ ) 1 / 2 ∑ n ≤ T t / γ ( θ n + 1 − E [ θ n + 1 ∣ F n ] ) , t ∈ [ 0 , 1 ] , \mathcal{P}^{\mathcal{T}}_t=\left(\frac{\gamma}{\mathcal{T}\overline{I}}\right)^{1/2}\sum_{n\leq\mathcal{T}t/\gamma}\left(\theta_{n+1}-\mathbb{E}[\theta_{n+1}\mid\mathcal{F}_n]\right),\qquad t\in[0,1], P t T = ( T I γ ) 1/2 n ≤ T t / γ ∑ ( θ n + 1 − E [ θ n + 1 ∣ F n ] ) , t ∈ [ 0 , 1 ] , denote the rescaled fluctuation of the mid price. Then every sequence T k → ∞ \mathcal{T}_k\to\infty T k → ∞ has a subsequence along which ( P T k , X T k ) (\mathcal{P}^{\mathcal{T}_k},\mathcal{X}^{\mathcal{T}_k}) ( P T k , X T k ) converges in law, and every limit satisfies
P t = ∫ 0 t X s ′ d W s , t ∈ [ 0 , 1 ] , \mathcal{P}_t=\int_0^t\sqrt{\mathcal{X}'_s}\,dW_s,\qquad t\in[0,1], P t = ∫ 0 t X s ′ d W s , t ∈ [ 0 , 1 ] , for a Brownian motion W W W and the derivative X ′ \mathcal{X}' X ′ of Theorem 5.3 .
In the limit, the squared volatility X ′ \mathcal{X}' X ′ of the mid price is thus as rough as fractional Brownian motion with the Hurst parameter H = α − 1 / 2 H=\alpha-1/2 H = α − 1/2 .
Proof. Given F n \mathcal{F}_n F n , the indicators ξ i ( n ) \xi_i(n) ξ i ( n ) are independent (§ 3.2 ), while Δ Z i ( n ) \Delta Z_i(n) Δ Z i ( n ) and Z ~ n σ \widetilde{Z}^{\sigma}_n Z n σ are F n \mathcal{F}_n F n -measurable (proof of Lemma 4.2 ). With (7) , we obtain
Var [ θ n + 1 ∣ F n ] = ∑ i ( c m i ∣ Δ Z i ( n ) ∣ ) 2 p k ( i ) ( n ) ( 1 − p k ( i ) ( n ) ) = Var [ I n ∣ F n ] . \operatorname{Var}[\theta_{n+1}\mid\mathcal{F}_n]=\sum_i\left(c\,m_i\lvert\Delta Z_i(n)\rvert\right)^2p_{k(i)}(n)\left(1-p_{k(i)}(n)\right)=\operatorname{Var}[I_n\mid\mathcal{F}_n]. Var [ θ n + 1 ∣ F n ] = i ∑ ( c m i ∣ Δ Z i ( n )∣ ) 2 p k ( i ) ( n ) ( 1 − p k ( i ) ( n ) ) = Var [ I n ∣ F n ] .
The fluctuation of θ n + 1 \theta_{n+1} θ n + 1 differs from I ~ n \widetilde{I}_n I n only in the signs σ i \sigma_i σ i of its terms. Lemmas 5.9 and 5.10 and the proof of Theorem 5.3 thus apply to P T \mathcal{P}^{\mathcal{T}} P T in place of Z T \mathcal{Z}^{\mathcal{T}} Z T , such that every limit P \mathcal{P} P is a continuous martingale with [ P , P ] = X [\mathcal{P},\mathcal{P}]=\mathcal{X} [ P , P ] = X . By [RY99, Chapter V, (3.8)], we obtain P = ∫ 0 ⋅ X s ′ d W s \mathcal{P}=\int_0^{\cdot}\sqrt{\mathcal{X}'_s}\,dW_s P = ∫ 0 ⋅ X s ′ d W s .
Numerical experiments and results
We test each step of our mechanism on a simulation of the protocol of § 2.6 over a finite time. The steps lead from the book (Proposition 3.10 ) to the activity (Corollary 4.6 ) and on to the mid price (Corollary 5.12 ). For each step we compare a quantity measured in the simulation with its value in the limit.
Setting and implementation
We measure every order by its distance x > 0 x>0 x > 0 from the mid price (§ 2.1 ). As a function of this distance, the volume of average order books is known empirically to peak away from the best price [10 , 28 PB03] . We therefore draw each target from the Gamma density x e − x / 1.5 / 1.5 2 xe^{-x/1.5}/1.5^2 x e − x /1.5 /1. 5 2 (Table B.1). This density vanishes at x = 0 x=0 x = 0 and peaks at x = 1.5 x=1.5 x = 1.5 .
The Dean–Kawasaki check
For the test function φ d ( x ) = e − ( x − d ) 2 / 2 \varphi_d(x)=e^{-(x-d)^2/2} φ d ( x ) = e − ( x − d ) 2 /2 at the depth d ∈ { 1 , 3 , 6 } d\in\{1,3,6\} d ∈ { 1 , 3 , 6 } we follow the volume of the asks around d d d ,
V n ( d ) = ∑ i : σ i = + 1 m i ∑ j = 1 L φ d ( x i , j ( n ) ) . V_n(d)=\sum_{i:\sigma_i=+1}m_i\sum_{j=1}^{L}\varphi_d(x_{i,j}(n)). V n ( d ) = i : σ i = + 1 ∑ m i j = 1 ∑ L φ d ( x i , j ( n )) .
By Proposition 3.10 , its change Δ n ( d ) = V n + 1 ( d ) − V n ( d ) \Delta_n(d)=V_{n+1}(d)-V_n(d) Δ n ( d ) = V n + 1 ( d ) − V n ( d ) has the conditional mean μ n ( d ) \mu_n(d) μ n ( d ) . By Lemma 3.1 , the orders of one desk move together, such that the conditional variance sums over desks rather than over orders,
s n 2 = ∑ i : σ i = + 1 p k ( i ) ( 1 − p k ( i ) ) ( m i ∑ j φ d ′ ( x i , j ) g i , j ) 2 , s_n^2=\sum_{i:\sigma_i=+1}p_{k(i)}(1-p_{k(i)})\left(m_i\sum_j\varphi'_d(x_{i,j})g_{i,j}\right)^2, s n 2 = i : σ i = + 1 ∑ p k ( i ) ( 1 − p k ( i ) ) ( m i j ∑ φ d ′ ( x i , j ) g i , j ) 2 ,
( s n i n d ) 2 = ∑ i : σ i = + 1 ∑ j p k ( i ) ( 1 − p k ( i ) ) ( m i φ d ′ ( x i , j ) g i , j ) 2 . \left(s_n^{\mathrm{ind}}\right)^2=\sum_{i:\sigma_i=+1}\sum_j p_{k(i)}(1-p_{k(i)})\left(m_i\varphi'_d(x_{i,j})g_{i,j}\right)^2. ( s n ind ) 2 = i : σ i = + 1 ∑ j ∑ p k ( i ) ( 1 − p k ( i ) ) ( m i φ d ′ ( x i , j ) g i , j ) 2 .
In the limit the slope of Δ n ( d ) \Delta_n(d) Δ n ( d ) on μ n ( d ) \mu_n(d) μ n ( d ) is one, and ( Δ n ( d ) − μ n ( d ) ) / s n (\Delta_n(d)-\mu_n(d))/s_n ( Δ n ( d ) − μ n ( d )) / s n has variance one.
The slope lies within 7% of one, and the variance with s n s_n s n within 1.5% of one (Figure 1 a). With independent orders the variance lies between 1.18 and 4.23.
Figure 1. Simulation against Proposition 3.10 and Corollary 4.6 . (a) Change Δ n ( d ) \Delta_n(d) Δ n ( d ) of the book against its conditional mean μ n ( d ) \mu_n(d) μ n ( d ) of Proposition 3.10 , both divided by the standard deviation of μ n ( d ) \mu_n(d) μ n ( d ) , in 15 bins of equal count over d ∈ { 1 , 3 , 6 } d \in\{1,3,6\} d ∈ { 1 , 3 , 6 } (L = 6 L = 6 L = 6 , Table 1 ). Bars are standard errors. (b) Response D ( s ) D(s) D ( s ) of the activity to Δ I = 4 ⋅ 10 − 4 \Delta I = 4 \cdot10^{-4} Δ I = 4 ⋅ 1 0 − 4 added at step n 0 n_0 n 0 , averaged over 600 pairs, against ( Δ I / κ ˉ ) R ( s ) (\Delta I/\bar{\kappa})R(s) ( Δ I / κ ˉ ) R ( s ) (Table 2 ).
Table 1. Change of the book against Proposition 3.10 (β κ ˉ = 0.5 \beta\bar{\kappa}=0.5 β κ ˉ = 0.5 , 10 5 10^5 1 0 5 steps).
Table 2. Measured over predicted response, D ( s ) / ( ( Δ I / κ ˉ ) R ( s ) ) D(s)/((\Delta I/\bar{\kappa})R(s)) D ( s ) / (( Δ I / κ ˉ ) R ( s )) , averaged over 600 pairs.
The memory kernel
To measure the response R R R , we add Δ I = 4 ⋅ 10 − 4 \Delta I=4\cdot10^{-4} Δ I = 4 ⋅ 1 0 − 4 to I n 0 I_{n_0} I n 0 at β κ ˉ = 0.1 \beta\bar{\kappa}=0.1 β κ ˉ = 0.1 and compare the activity with that of the same run without this addition. All clusters share the target density, such that we use the measured κ ˉ \bar{\kappa} κ ˉ for every cluster. By Proposition 4.3 , the activity responds linearly as long as no p k p_k p k reaches one (Table B.3). The difference D ( s ) D(s) D ( s ) of the two activities at s = n − n 0 s=n-n_0 s = n − n 0 then solves the renewal equation
D ( s ) = Δ I κ ˉ ϕ ( s ) + ∑ r = 1 s − 1 ϕ ( s − r ) D ( r ) , ϕ ( s ) = β κ ˉ ∑ k w k ω k ( s ) , D(s)=\frac{\Delta I}{\bar{\kappa}}\phi(s)+\sum_{r=1}^{s-1}\phi(s-r)D(r),\qquad\phi(s)=\beta\bar{\kappa}\sum_k w_k\omega_k(s), D ( s ) = κ ˉ Δ I ϕ ( s ) + r = 1 ∑ s − 1 ϕ ( s − r ) D ( r ) , ϕ ( s ) = β κ ˉ k ∑ w k ω k ( s ) ,
such that D = ( Δ I / κ ˉ ) R D=(\Delta I/\bar{\kappa})R D = ( Δ I / κ ˉ ) R . As both runs of a pair share their random numbers, the ratio of the measured to the predicted response is one up to the sampling error of 600 pairs.
The ratio stays within 2% of one over three decades of lags (Figure 1 b). The horizons of Table B.1 grow by the factor Ω = 1.49 \Omega=1.49 Ω = 1.49 . By Remark 4.12 , the kernel then deviates from a power law by at most
b ( Ω , α ) = 2 I ∞ ∑ m ≥ 1 ∣ J m ∣ = 2 Γ ( 1 + α ) ∑ m ≥ 1 ∣ Γ ( 1 + α + 2 π i m ln Ω ) ∣ , b(\Omega,\alpha)=\frac{2}{I_\infty}\sum_{m\geq1}|J_m|=\frac{2}{\Gamma(1+\alpha)}\sum_{m\geq1}\left|\Gamma\left(1+\alpha+\frac{2\pi i m}{\ln\Omega}\right)\right|, b ( Ω , α ) = I ∞ 2 m ≥ 1 ∑ ∣ J m ∣ = Γ ( 1 + α ) 2 m ≥ 1 ∑ Γ ( 1 + α + ln Ω 2 π im ) ,
which we evaluate numerically.
Table 3. Bound b ( Ω , α ) b(\Omega,\alpha) b ( Ω , α ) on the deviation of the kernel from a power law.
For Ω = 1.49 \Omega= 1.49 Ω = 1.49 the bound lies below 3.1 ⋅ 10 − 9 3.1 \cdot10^{-9} 3.1 ⋅ 1 0 − 9 , such that for lags well inside the range of horizons the kernel is a power law to this accuracy.
Roughness
We compute two series over bins b b b of 50 steps,
A b = ∑ n ∈ b E [ A n ∣ F n ] , V b = ∑ n ∈ b Var [ θ n + 1 ∣ F n ] . A_b = \sum_{n \in b} \mathbb{E}[A_n \mid\mathcal{F}_n], \qquad V_b = \sum_{n \in b} \operatorname{Var}[\theta_{n+1} \mid\mathcal{F}_n]. A b = n ∈ b ∑ E [ A n ∣ F n ] , V b = n ∈ b ∑ Var [ θ n + 1 ∣ F n ] .
We use the conditional mean of the activity because the actions of one step add noise without memory to the variogram. For each series Y ∈ { A , V } Y \in\{A,V\} Y ∈ { A , V } we estimate the Hurst parameter H H H from the variogram
m Y ( Δ ) = E [ ( ln Y b + Δ − ln Y b ) 2 ] ∝ Δ 2 H m_Y(\Delta) = \mathbb{E}\left[(\ln Y_{b+\Delta} - \ln Y_b)^2\right] \propto\Delta^{2H} m Y ( Δ ) = E [ ( ln Y b + Δ − ln Y b ) 2 ] ∝ Δ 2 H
over the lags of Table 4 (Appendix § 10 ). By Theorem 5.3 , the activity has H = α − 1 2 H = \alpha- \frac{1}{2} H = α − 2 1 as T → ∞ \mathcal{T} \to\infty T → ∞ . By the proof of Corollary 5.12 , V b V_b V b equals ∑ n ∈ b Var [ I n ∣ F n ] \sum_{n \in b} \operatorname{Var}[I_n \mid\mathcal{F}_n] ∑ n ∈ b Var [ I n ∣ F n ] . We set
Table 4. Hurst parameters of the activity and of the squared volatility of the mid price (bins of 50 steps, Table B.2).
q ˉ b = ∑ n ∈ b Var [ I n ∣ F n ] ∑ n ∈ b E [ I n ∣ F n ] . \bar{q}_b = \frac{\sum_{n \in b} \operatorname{Var}[I_n \mid\mathcal{F}_n]}{\sum_{n \in b} \mathbb{E}[I_n \mid\mathcal{F}_n]}. q ˉ b = ∑ n ∈ b E [ I n ∣ F n ] ∑ n ∈ b Var [ I n ∣ F n ] .
With E [ I n ∣ F n ] = κ ˉ E [ A n ∣ F n ] \mathbb{E}[I_n \mid\mathcal{F}_n] = \bar{\kappa}\mathbb{E}[A_n \mid\mathcal{F}_n] E [ I n ∣ F n ] = κ ˉ E [ A n ∣ F n ] (Proposition 4.5 ), we obtain
ln V b = ln κ ˉ + ln A b + ln q ˉ b . \ln V_b = \ln\bar{\kappa} + \ln A_b + \ln\bar{q}_b. ln V b = ln κ ˉ + ln A b + ln q ˉ b .
By Lemma 5.9 , q ˉ b \bar{q}_b q ˉ b is close to the constant ν ˉ \bar{\nu} ν ˉ . Both series thus have the same variogram up to the variation of q ˉ b \bar{q}_b q ˉ b , and hence the same H H H .
The two estimates agree within 0.016. For α = 0.7 \alpha= 0.7 α = 0.7 both equal α − 1 2 \alpha- \frac{1}{2} α − 2 1 . For α = 0.6 \alpha= 0.6 α = 0.6 both exceed it by 0.05 at our finite T \mathcal{T} T . With a single cluster the kernel has the finite mean ∑ s > 1 ω ( s ) = T \sum_{s>1}\omega(s) = \mathcal{T} ∑ s > 1 ω ( s ) = T . By [40 Theorem 2.2] , the activity then tends to the solution of d X t = a ( μ − X t ) d t + σ X t d B t dX_t = a(\mu- X_t)\,dt + \sigma\sqrt{X_t}\,dB_t d X t = a ( μ − X t ) d t + σ X t d B t , which has H = 1 2 H = \frac{1}{2} H = 2 1 by Lemma B.1 . Both estimates lie within 0.03 of this limit.
The role of the book
The squared volatility of the mid price in § 6.4 is the conditional variance of one function of the book. For the changes of the book at the depths d d d of § 6.2 we sum the same variance over each bin,
S b ( d ) = ∑ n ∈ b ∑ i : σ i = + 1 p k ( n ) ( 1 − p k ( n ) ) ( m i ∑ j [ φ d ( x i , j n e w ) − φ d ( x i , j ) ] ) 2 S_b(d) = \sum_{n \in b} \sum_{i:\,\sigma_i=+1} p_k(n)\left(1-p_k(n)\right)\left(m_i \sum_j \left[\varphi_d\left(x_{i,j}^{\mathrm{new}}\right)-\varphi_d\left(x_{i,j}\right)\right]\right)^2 S b ( d ) = n ∈ b ∑ i : σ i = + 1 ∑ p k ( n ) ( 1 − p k ( n ) ) ( m i j ∑ [ φ d ( x i , j new ) − φ d ( x i , j ) ] ) 2
where k k k is the cluster of desk i i i . With r b ( d ) = S b ( d ) / A b r_b(d) = S_b(d)/A_b r b ( d ) = S b ( d ) / A b we obtain
ln S b ( d ) = ln A b + ln r b ( d ) . \ln S_b(d) = \ln A_b + \ln r_b(d). ln S b ( d ) = ln A b + ln r b ( d ) .
The book thus has the activity’s H H H when the variogram of ln r b ( d ) \ln r_b(d) ln r b ( d ) is small against that of ln A b \ln A_b ln A b . By Lemma 4.4 , the book recovers from a shift within about 1 / p k 1/p_k 1/ p k steps, while the memory of cluster k k k lasts about T k T_k T k steps. With the mean activity of 0.05 (Proposition 4.8 ), the product T k p k T_kp_k T k p k lies between 50 and 250 for horizons in the range of lags of Table 4 . A single cluster with T = 200 T = 200 T = 200 instead has T p k ≈ 10 Tp_k \approx10 T p k ≈ 10 . The book therefore recovers its shape within the memory of the desks for 24 clusters but not for a single cluster.
For 24 clusters the variogram of ln r b ( d ) \ln r_b(d) ln r b ( d ) is at most 0.08 times that of ln A b \ln A_b ln A b . At every depth the book then has the activity’s H H H within 0.016. With a single cluster the ratio of the two variograms lies between 0.5 and 4.1. The H H H of the book then falls to between 0.23 and 0.29.
The norm ∥ ϕ ∥ = β κ ˉ \lVert\phi\rVert= \beta\bar{\kappa} ∥ ϕ ∥ = β κ ˉ of Proposition 4.7 depends on the book through κ ˉ = ∑ l κ l w l \bar{\kappa} = \sum_l \kappa_l w_l κ ˉ = ∑ l κ l w l . By the definition of κ l \kappa_l κ l before Lemma 4.2 , we have
κ ˉ = c ∑ i m i ∣ Δ Z i ( n ) ∣ ≈ c γ ℓ ∑ i m i ∣ ∑ j e − x i , j / ℓ ( x i , j − x i , j ∗ ) ∣ \bar{\kappa} = c \sum_i m_i \lvert\Delta Z_i(n) \rvert\approx\frac{c\gamma}{\ell} \sum_i m_i \left\lvert\sum_j e^{-x_{i,j}/\ell}(x_{i,j} - x^*_{i,j}) \right\rvert κ ˉ = c i ∑ m i ∣ Δ Z i ( n )∣ ≈ ℓ c γ i ∑ m i j ∑ e − x i , j / ℓ ( x i , j − x i , j ∗ )
where the second step keeps the first order in γ \gamma γ and drops the trend P k P_k P k . Orders further from the mid price enter this sum with a smaller factor e − x i , j / ℓ e^{-x_{i,j}/\ell} e − x i , j / ℓ but with larger gaps ∣ x i , j − x i , j ∗ ∣ \lvert x_{i,j} - x^*_{i,j} \rvert ∣ x i , j − x i , j ∗ ∣ (Table 6 ). The norm can therefore peak at an intermediate distance of the targets. We test this prediction at a fixed β \beta β by multiplying every target by s / 1.5 s/1.5 s /1.5 . The density of Table B.1 then becomes x e − x / s / s 2 x e^{-x/s}/s^2 x e − x / s / s 2 , which peaks at x = s x=s x = s .
Table 6 . Norm of the kernel as the targets move (24 clusters, α = 0.6 \alpha= 0.6 α = 0.6 , β \beta β fixed at β κ ˉ = 0.5 \beta\bar{\kappa} = 0.5 β κ ˉ = 0.5 for s = 1.5 s = 1.5 s = 1.5 ). Here g ˉ = ∑ i , j m i ∣ x i , j − x i , j ∗ ∣ / ∑ i , j m i \bar{g} = \sum_{i,j}m_i\lvert x_{i,j} - x^*_{i,j}\rvert/\sum_{i,j}m_i g ˉ = ∑ i , j m i ∣ x i , j − x i , j ∗ ∣ / ∑ i , j m i is the mean gap, and both Z + + Z − Z^+ + Z^- Z + + Z − and g ˉ \bar{g} g ˉ are averages along the run.
Table 6. Norm of the kernel as the targets move (24 clusters, α = 0.6 \alpha= 0.6 α = 0.6 , β \beta β fixed at β κ ˉ = 0.5 \beta\bar{\kappa} = 0.5 β κ ˉ = 0.5 for s = 1.5 s = 1.5 s = 1.5 ). Here g ˉ = ∑ i , j m i ∣ x i , j − x i , j ∗ ∣ / ∑ i , j m i \bar{g} = \sum_{i,j}m_i\lvert x_{i,j} - x^*_{i,j}\rvert/\sum_{i,j}m_i g ˉ = ∑ i , j m i ∣ x i , j − x i , j ∗ ∣ / ∑ i , j m i is the mean gap, and both Z + + Z − Z^+ + Z^- Z + + Z − and g ˉ \bar{g} g ˉ are averages along the run.
The norm peaks at s = 1 s = 1 s = 1 and falls on both sides. At every s s s the first-order value lies 3–5% below the measured norm. Between the tenths of each run κ ˉ \bar{\kappa} κ ˉ varies by at most 0.23%, which supports replacing κ l ( n ) \kappa_l(n) κ l ( n ) by its average in § 4.1 .
Evidence from Binance
We test our mechanism on a market whose book market makers change mainly by moving orders. The exchange Binance offers such a market in its perpetual futures of BTC, ETH and SOL against USDT. Trades there rarely reach 1% from the mid price (Table B.4). Market makers thus change the volume beyond 1% mainly by placing and cancelling orders. Under price-time priority [GPW+13], an order that a market maker cancels and places again loses only its place in the queue. As queues are short at a small tick [3 ] , this place is worth little at the tick of 10 − 6 10^{-6} 1 0 − 6 to 10 − 4 10^{-4} 1 0 − 4 of the price of these markets (Table B.4). The public data of Binance [7 ] contain the taker orders of these perpetual futures and of the spot markets of the same coins. As the data contain no actions of market makers, we count the taker orders N b N_b N b in each bin b b b as the activity. In snapshot n n n of a book, V n ( d ) V_n(d) V n ( d ) denotes the volume resting between d % d\% d % and ( d + 1 ) % (d+1)\% ( d + 1 ) % from the mid price. The move of the mid price up to snapshot n + 1 n+1 n + 1 changes V n ( d ) V_n(d) V n ( d ) by Δ n m i d ( d ) \Delta_n^{\mathrm{mid}}(d) Δ n mid ( d ) when every order keeps its price (Algorithm B.1). With the log return r i r_i r i over minute i i i , we measure
R V b = ∑ i ∈ b r i 2 , ς = ∑ n , d Δ V n ( d ) Δ n m i d ( d ) ∑ n , d Δ n m i d ( d ) 2 , Q b = ∑ n ∈ b ∑ d = 1 4 e n ( d ) 2 ∑ n ∈ b ( t n + 1 − t n ) \mathrm{RV}_b=\sum_{i\in b}r_i^2,\qquad\varsigma=\frac{\sum_{n,d}\Delta V_n(d)\Delta_n^{\mathrm{mid}}(d)}{\sum_{n,d}\Delta_n^{\mathrm{mid}}(d)^2},\qquad Q_b=\frac{\sum_{n\in b}\sum_{d=1}^{4}e_n(d)^2}{\sum_{n\in b}(t_{n+1}-t_n)} RV b = i ∈ b ∑ r i 2 , ς = ∑ n , d Δ n mid ( d ) 2 ∑ n , d Δ V n ( d ) Δ n mid ( d ) , Q b = ∑ n ∈ b ( t n + 1 − t n ) ∑ n ∈ b ∑ d = 1 4 e n ( d ) 2
with Δ V n ( d ) = V n + 1 ( d ) − V n ( d ) \Delta V_n(d)=V_{n+1}(d)-V_n(d) Δ V n ( d ) = V n + 1 ( d ) − V n ( d ) and e n ( d ) = Δ V n ( d ) − ς Δ n m i d ( d ) e_n(d)=\Delta V_n(d)-\varsigma\Delta_n^{\mathrm{mid}}(d) e n ( d ) = Δ V n ( d ) − ς Δ n mid ( d ) . A book whose orders all keep their prices has ς = 1 \varsigma=1 ς = 1 . With the Hurst parameter H ( Y ) H(Y) H ( Y ) of a series Y Y Y (Appendix § 10 ), we then test
H ( N b ) < 1 2 six markets x r e f ‘ T h e o r e m 5.3 ‘ , ∥ ϕ ∥ ≈ 1 , ϕ ( s ) ∝ s − 1 − α six markets x r e f ‘ S e c t i o n 4.3 ‘ , H ( R V b ) = H ( N b ) six markets x r e f ‘ C o r o l l a r y 5.12 ‘ , Corr ( Δ V n ( d ) , Δ V n ( d + 1 ) ) < 0 perpetual futures x r e f ‘ P r o p o s i t i o n 3.10 ‘ , ς < 1 perpetual futures x r e f ‘ L e m m a 4.4 ‘ , H ( Q b ) = H ( N b ) perpetual futures x r e f ‘ S e c t i o n 6.5 ‘ . \begin{aligned}
H(N_b)&<\frac{1}{2} && \text{six markets} && {xref}`Theorem 5.3`,\\
\lVert\phi\rVert&\approx1,\quad\phi(s)\propto s^{-1-\alpha} && \text{six markets} && {xref}`Section 4.3`,\\
H(\mathrm{RV}_b)&=H(N_b) && \text{six markets} && {xref}`Corollary 5.12`,\\
\operatorname{Corr}(\Delta V_n(d),\Delta V_n(d+1))&<0 && \text{perpetual futures} && {xref}`Proposition 3.10`,\\
\varsigma&<1 && \text{perpetual futures} && {xref}`Lemma 4.4`,\\
H(Q_b)&=H(N_b) && \text{perpetual futures} && {xref}`Section 6.5`.
\end{aligned} H ( N b ) ∥ ϕ ∥ H ( RV b ) Corr ( Δ V n ( d ) , Δ V n ( d + 1 )) ς H ( Q b ) < 2 1 ≈ 1 , ϕ ( s ) ∝ s − 1 − α = H ( N b ) < 0 < 1 = H ( N b ) six markets six markets six markets perpetual futures perpetual futures perpetual futures x r e f ‘ T h eor e m 5.3‘ , x r e f ‘ S ec t i o n 4.3‘ , x r e f ‘ C or o l l a r y 5.12‘ , x r e f ‘ P r o p os i t i o n 3.10‘ , x r e f ‘ L e mma 4.4‘ , x r e f ‘ S ec t i o n 6.5‘.
Data
The relations above need the activity over time scales from one second to one day. Over the 180 days from 10 April to 6 October 2026, we therefore count per second the aggregated trades of each market. Binance merges into one aggregated trade the fills on one side at one price within 100 ms [7 ] . As the shortest lag of 15 minutes needs several bins, N b N_b N b counts these trades in bins b b b of 5 minutes. The sampling noise of n n n one-minute returns adds to the variogram of ln R V b \ln\mathrm{RV}_b ln RV b a constant of 2 ψ 1 ( n / 2 ) 2\psi_1(n/2) 2 ψ 1 ( n /2 ) for Gaussian returns, where ψ 1 \psi_1 ψ 1 is the trigamma function (Appendix § 10 ). Measured on these markets, this constant falls from 0.38–0.63 at bins of 15 minutes to 0.12–0.15 at bins of 60 minutes. We therefore compare H ( R V b ) H(\mathrm{RV}_b) H ( RV b ) with H ( N b ) H(N_b) H ( N b ) at bins of 15, 30 and 60 minutes. For the perpetual futures, Binance publishes a snapshot of the book about every 30 seconds with the cumulative volume within 0.2% and within 1% to 5% of the mid price on each side [7 ] .
The activity of such markets is known empirically to follow a daily and a weekly pattern [4 , 36 ] . A pattern s b s_b s b that repeats every day or every week adds to the variogram of ln N b = s b + y b \ln N_b=s_b+y_b ln N b = s b + y b a second term,
m ln N ( Δ ) = m y ( Δ ) + 1 B ∑ b = 1 B ( s b + Δ − s b ) 2 . m_{\ln N}(\Delta)=m_y(\Delta)+\frac{1}{B}\sum_{b=1}^{B}(s_{b+\Delta}-s_b)^2. m l n N ( Δ ) = m y ( Δ ) + B 1 b = 1 ∑ B ( s b + Δ − s b ) 2 .
This term grows with Δ \Delta Δ up to half a day and raises the estimated H H H . Removing the pattern thus lowers H H H . As estimation error alone can produce H < 1 2 H<\frac{1}{2} H < 2 1 [12 ] , we also estimate H ( N b ) H(N_b) H ( N b ) without any removal. It then lies between 0.10 and 0.14 (Table A.2). As these markets trade around the clock, s b s_b s b has no jump at an open or a close. With the bin τ b ∈ { 0 , … , 287 } \tau_b\in\{0,\ldots,287\} τ b ∈ { 0 , … , 287 } of the day and the weekday δ b ∈ { 1 , … , 7 } \delta_b\in\{1,\ldots,7\} δ b ∈ { 1 , … , 7 } of bin b b b , we therefore fit by least squares
ln N b = ∑ k = 1 6 ( a k cos 2 π k τ b 288 + c k sin 2 π k τ b 288 ) + w δ b + y b . \ln N_b=\sum_{k=1}^{6}\left(a_k\cos\frac{2\pi k\tau_b}{288}+c_k\sin\frac{2\pi k\tau_b}{288}\right)+w_{\delta_b}+y_b. ln N b = k = 1 ∑ 6 ( a k cos 288 2 π k τ b + c k sin 288 2 π k τ b ) + w δ b + y b .
We estimate the Hurst parameter of the activity from the remainder y b y_b y b (Appendix § 10 ).
Activity and volatility
We estimate H ( N b ) H(N_b) H ( N b ) from the variogram of y b y_b y b over lags of 15 minutes to one day, in each of the six blocks of 30 days (Appendix § 10 ). The kernel of our desks in § 4.3 has ω k ( s ) = T k − 1 ( 1 − T k − 1 ) s − 1 \omega_k(s) = T_k^{-1}(1-T_k^{-1})^{s-1} ω k ( s ) = T k − 1 ( 1 − T k − 1 ) s − 1 , which becomes e − s / T k / T k e^{-s/T_k}/T_k e − s / T k / T k for horizons T k T_k T k in seconds. With 80 horizons T k = c ( 10 7 s / c ) ( k − 1 ) / 79 T_k = c(10^7\mathrm{s}/c)^{(k-1)/79} T k = c ( 1 0 7 s / c ) ( k − 1 ) /79 , we fit to the spectrum of the counts per second over time scales of 1 minute to 4 hours (Appendix § 10 ).
ϕ d e s k ( s ) = a ∑ k = 1 80 w k e − s / T k T k , w k ∝ T k − α , a = β κ ˉ . \phi_{\mathrm{desk}}(s) = a\sum_{k=1}^{80}w_k\frac{e^{-s/T_k}}{T_k},\qquad w_k\propto T_k^{-\alpha},\qquad a=\beta\bar{\kappa}. ϕ desk ( s ) = a k = 1 ∑ 80 w k T k e − s / T k , w k ∝ T k − α , a = β κ ˉ .
We also fit ϕ p o w ( s ) = a α c α ( s + c ) − 1 − α \phi_{\mathrm{pow}}(s)=a\alpha c^\alpha(s+c)^{-1-\alpha} ϕ pow ( s ) = a α c α ( s + c ) − 1 − α , which has the same norm and tail but another shape at time scales near c c c .
In every market H ( N b ) H(N_b) H ( N b ) lies between 0.07 and 0.09 (Figure 2 a). The 95% lower bound of the norm a a a is at least 0.94. The order flow thus lies close to the threshold of Theorem 5.3 . The fitted tails lie between 0.39 and 0.50, where the limit H = α − 1 2 H=\alpha-\frac{1}{2} H = α − 2 1 of that theorem is not positive. At distances 1 − a = T − α 1-a=T^{-\alpha} 1 − a = T − α with T T T of 10 5 10^5 1 0 5 or 10 6 s 10^6\mathrm{s} 1 0 6 s , simulated desks with these tails give values of H H H that the data exceed by 0.016 and 0.009 on average, with 95% intervals of [− 0.005 -0.005 − 0.005 , 0.036] and [− 0.012 -0.012 − 0.012 , 0.030] (Table 7 ). At T = 10 4 s T=10^4\mathrm{s} T = 1 0 4 s the desks fall short of the data by 0.039 [0.023, 0.055]. The data thus agree with desks whose memory reaches beyond the longest lag of one day. As both kernels share their tail, the gain of ϕ d e s k \phi_{\mathrm{desk}} ϕ desk in three markets reflects its shape near c c c . A similar phenomenon has been observed in the term structure of the at-the-money skew of equity indices [29 ] . There, a power law fits beyond short maturities, while shapes that resemble it but stay finite fit better at short ones.
Table 7. Activity and kernel of the six markets, 10 April to 6 October 2026. The gain is the log-likelihood of ϕ d e s k \phi_{\mathrm{desk}} ϕ desk minus that of ϕ p o w \phi_{\mathrm{pow}} ϕ pow . The last column gives H H H of the desks of Table B.6 at the fitted α \alpha α , for T = 10 5 T=10^5 T = 1 0 5 and 10 6 s 10^6\mathrm{s} 1 0 6 s .
Figure 2. Activity and kernel of the six markets, 10 April to 6 October 2026. (a) H ( N b ) H(N_b) H ( N b ) in each block of 30 days, with the mean over the blocks. (b) Periodogram I ( ν ) I(\nu) I ( ν ) of the counts per second, scaled to 1 at 1 minute, with the fits of ϕ d e s k \phi_{\mathrm{desk}} ϕ desk over the shaded range (Appendix § 10 .§ 4 ).
By Corollary 5.12 the squared volatility of the mid price has the Hurst parameter of the activity. We estimate this squared volatility by R V b RV_b R V b , whose sampling noise adds a constant η \eta η to the variogram of ln R V b \ln RV_b ln R V b (§ 7.1 ). To measure η \eta η , we split the n n n minutes of each bin into its odd and its even minutes,
R V b o d d = ∑ i ∈ b , i o d d r i 2 , R V b e v e n = ∑ i ∈ b , i e v e n r i 2 . RV_b^{\mathrm{odd}} = \sum_{\substack{i\in b,\,i\ \mathrm{odd}}} r_i^2, \qquad RV_b^{\mathrm{even}} = \sum_{\substack{i\in b,\,i\ \mathrm{even}}} r_i^2. R V b odd = i ∈ b , i odd ∑ r i 2 , R V b even = i ∈ b , i even ∑ r i 2 .
As both halves share the volatility of the bin, the difference of their logarithms contains only sampling noise. With the trigamma function ψ 1 \psi_1 ψ 1 of § 7.1 , the variance of this difference gives
η ^ = 2 ψ 1 ( n 2 ) Var b [ ln R V b o d d − ln R V b e v e n ] ψ 1 ( ⌈ n / 2 ⌉ / 2 ) + ψ 1 ( ⌊ n / 2 ⌋ / 2 ) . \widehat{\eta}=2\psi_1\left(\frac{n}{2}\right)\frac{\operatorname{Var}_b\left[\ln RV_b^{\mathrm{odd}}-\ln RV_b^{\mathrm{even}}\right]}{\psi_1\left(\lceil n/2\rceil/2\right)+\psi_1\left(\lfloor n/2\rfloor/2\right)}. η = 2 ψ 1 ( 2 n ) ψ 1 ( ⌈ n /2 ⌉ /2 ) + ψ 1 ( ⌊ n /2 ⌋ /2 ) Var b [ ln R V b odd − ln R V b even ] .
We subtract η ^ \widehat{\eta} η from the variogram of ln R V b \ln RV_b ln R V b , and estimate H ( R V b ) H(RV_b) H ( R V b ) (Appendix § 10 ).
At every bin length H ( R V b ) H(RV_b) H ( R V b ) lies within 0.018 of H ( N b ) H(N_b) H ( N b ) (Table 8 ). The logarithms of the two series have correlations of 0.83 to 0.95. From bins of 15 to 60 minutes η ^ \widehat{\eta} η falls by a factor of four without changing this agreement. Averaging within a bin raises the fitted H H H of a rough series. A log-intensity with H = 0.08 H = 0.08 H = 0.08 at one-minute resolution gives 0.16, 0.19 and 0.23 at bins of 15, 30 and 60 minutes (Table A.3).
Table 8. Hurst parameters of the activity and of the realised variance in the six markets, 10 April to 6 October 2026, at lags of 1 hour to 1 day. Here N b N_b N b counts the trades of the one-minute candles, and the pattern removed is the mean of each bin of the day and of each weekday.
The order book
As trades almost never reach 1% from the mid price (Table B.4), we measure the books of the perpetual futures between 1% and 5%. A move of the mid price changes the distance of every order that keeps its price. By Lemma 4.4 , market makers move their orders back towards their distances. Let the mid price move at a constant speed v v v over an interval of length u u u . Let δ ( t ) \delta(t) δ ( t ) denote the displacement of an order from its distance at the start of the interval. If market makers move the order back at the rate 1 / τ 1/\tau 1/ τ , then
d δ d t = v − δ τ , δ ( 0 ) = 0 ⟹ δ ( u ) = v τ ( 1 − e − u / τ ) . \frac{d\delta}{dt} = v - \frac{\delta}{\tau}, \qquad\delta(0) = 0 \qquad\Longrightarrow\qquad\delta(u) = v\tau\left(1 - e^{-u/\tau}\right). d t d δ = v − τ δ , δ ( 0 ) = 0 ⟹ δ ( u ) = v τ ( 1 − e − u / τ ) .
As Δ V n ( d ) \Delta V_n(d) Δ V n ( d ) is linear in the displacements of the orders for small moves, the slope ς \varsigma ς equals the ratio of δ ( u ) \delta(u) δ ( u ) to the displacement v u vu v u of an order that keeps its price,
ς = δ ( u ) v u = 1 − e − u / τ u / τ . \varsigma= \frac{\delta(u)}{vu} = \frac{1-e^{-u/\tau}}{u/\tau}. ς = v u δ ( u ) = u / τ 1 − e − u / τ .
With u u u the median time between snapshots, we solve this relation for the time τ \tau τ of re-centring. Market makers change the volume e n ( d ) = Δ V n ( d ) − ς Δ n m i d ( d ) e_n(d) = \Delta V_n(d) - \varsigma\Delta_n^{\mathrm{mid}}(d) e n ( d ) = Δ V n ( d ) − ς Δ n mid ( d ) that remains after the move of the mid price by placing, cancelling and moving orders. A moved order leaves one interval for its neighbour and thus changes the two volumes in opposite directions, whereas orders placed and cancelled independently leave them uncorrelated. As the size of the changes varies strongly over a day, we correlate them within blocks B B B of 10 consecutive pairs of snapshots. Averaging over the three neighbouring pairs of the four intervals and taking the median over blocks and both sides of the book, we obtain
ρ = median B , ± 1 3 ∑ d = 1 3 Corr B ( e n ( d ) , e n ( d + 1 ) ) . \rho= \underset{B,\pm}{\operatorname{median}}\ \frac{1}{3}\sum_{d=1}^{3}\operatorname{Corr}_{B}\left(e_n(d),e_n(d+1)\right). ρ = B , ± median 3 1 d = 1 ∑ 3 Corr B ( e n ( d ) , e n ( d + 1 ) ) .
This statistic is zero for independent orders and negative for moved orders. Simulated books of our desks, passed through the same four intervals, give its value in our model (Table B.7). On pairs of snapshots whose move μ n \mu_n μ n of the mid price stays below 0.02% (Algorithm B.1), ρ \rho ρ needs no estimate of ς \varsigma ς .
By § 6.5 , the changes of the book have the Hurst parameter of the activity when the book recovers its shape within the memory of the desks. The sum of the squared remainders e n ( d ) e_n(d) e n ( d ) over a bin estimates the conditional variance S b ( d ) S_b(d) S b ( d ) of § 6.5 . As the time between snapshots varies, we divide this sum by the time that it covers, which gives Q b Q_b Q b . The odd and the even pairs of snapshots in a bin give two such ratios Q b o d d Q_b^{\mathrm{odd}} Q b odd and Q b e v e n Q_b^{\mathrm{even}} Q b even with independent sampling noise. As for RV, half the variance of ln Q b o d d − ln Q b e v e n \ln Q_b^{\mathrm{odd}}-\ln Q_b^{\mathrm{even}} ln Q b odd − ln Q b even estimates this noise in the variogram of ln Q b \ln Q_b ln Q b (Appendix § 10 ). We subtract it and remove the pattern of § 7.1 before we estimate H ( Q b ) H(Q_b) H ( Q b ) in bins of 5 minutes over lags of 15 minutes to one day. The time τ \tau τ of re-centring and the memory of the activity then decide whether § 6.5 predicts H ( Q b ) = H ( N b ) H(Q_b)=H(N_b) H ( Q b ) = H ( N b ) for these markets.
In all three books ς \varsigma ς lies between 0.37 and 0.53, such that market makers move their orders back within 12 to 21 seconds (Table 9 ). The orders that they move change neighbouring intervals in opposite directions, with ρ \rho ρ between − 0.30 -0.30 − 0.30 and − 0.16 -0.16 − 0.16 on all pairs and on quiet pairs. This range contains the values − 0.18 -0.18 − 0.18 to − 0.17 -0.17 − 0.17 of the books of our desks (Table B.7). As the books recover within seconds and the activity remembers beyond one day (§ 7.2 ), § 6.5 predicts H ( Q b ) = H ( N b ) H(Q_b) = H(N_b) H ( Q b ) = H ( N b ) . SOL meets this equality within 0.008, whereas for BTC and ETH H ( Q b ) H(Q_b) H ( Q b ) lies 0.03 and 0.04 below H ( N b ) H(N_b) H ( N b ) . The size of the changes of these books also follows the activity less closely than in our model, with correlations of 0.45 to 0.56 between the logarithms against 0.98 and 0.99 in Table 5 .
Table 5. Hurst parameters of the activity and of the changes of the book (bins of 50 steps, lags of Table 4 ). Each range runs over the depths d ∈ { 1 , 3 , 6 } d \in\{1,3,6\} d ∈ { 1 , 3 , 6 } and, in the third row, also over the shortest and the longest lag.
Table 5. Hurst parameters of the activity and of the changes of the book (bins of 50 steps, lags of Table 4). Each range runs over the depths d ∈ { 1 , 3 , 6 } d \in\{1,3,6\} d ∈ { 1 , 3 , 6 } and, in the third row, also over the shortest and the longest lag.
Table 9. The books of the perpetual futures, 7 September to 6 October 2026. Snapshots come every u = 30 u = 30 u = 30 s, and the mid price is taken from the one-second candles of the spot market. H H H is estimated at bins of 5 minutes over lags of 15 minutes to one day.
Conclusion
The rough volatility observed in many markets [8 , 31 , 49 Tak20] has so far been explained by traders who split large orders or by the absence of arbitrage [22 , 42 ] . Such explanations leave out the market makers who provide a large share of the liquidity in electronic markets [39 , 51 ] . We therefore studied desks of market makers that move their orders by gradient steps at times set by new information over their own horizons. The volume of their book follows a conservative stochastic partial differential equation whose fluctuations make the order flow self-exciting with a kernel ϕ \phi ϕ . The shape of the book sets the norm of ϕ \phi ϕ , such that close to the threshold the activity and the squared volatility of the mid price become rough with H = α − 1 2 H = \alpha- \frac{1}{2} H = α − 2 1 . For the book itself the mechanism predicts ς < 1 \varsigma< 1 ς < 1 , ρ < 0 \rho< 0 ρ < 0 and H ( Q b ) = H ( N b ) H(Q_b) = H(N_b) H ( Q b ) = H ( N b ) (§ 7 ).
Simulations over a finite time reproduce each of these steps and give roughness only when the desks sample information over distinct horizons (§ 6 ). On six markets of Binance the order flow and the volatility are rough with H H H between 0.07 and 0.09 and within 0.02 of each other. In the books of these markets, market makers move orders back within 12 to 21 seconds and move volume between neighbouring depths. These data agree with our desks at a finite distance from the threshold (§ 7.2 ).
We leave extensions such as a limit theorem for tails α ≤ 1 2 \alpha\le\frac{1}{2} α ≤ 2 1 , tests on data of single orders and markets with lower liquidity to future work.
Data and code availability
Code and processed data are available at https://github.com/rsn870/SGD_lob_rough_vol . Raw Binance data are public at https://data.binance.vision .
Estimators
Hurst parameter
Every Hurst parameter in § 6 and § 7 comes from the variogram of a logarithm. For a positive series Y b Y_b Y b with b = 1 , … , B b=1,\ldots,B b = 1 , … , B , let y b y_b y b denote ln Y b \ln Y_b ln Y b after the removal of the pattern of § 7.1 . In § 6 the simulated series have no pattern, and y b = ln Y b y_b=\ln Y_b y b = ln Y b . With the noise constant η \eta η of the series and a geometric grid L L L of lags, we compute
m ( Δ ) = 1 B − Δ ∑ b = 1 B − Δ ( y b + Δ − y b ) 2 , Δ ∈ L , m(\Delta)=\frac{1}{B-\Delta}\sum_{b=1}^{B-\Delta}(y_{b+\Delta}-y_b)^2,\qquad\Delta\in L, m ( Δ ) = B − Δ 1 b = 1 ∑ B − Δ ( y b + Δ − y b ) 2 , Δ ∈ L ,
( H ^ , c ^ ) = arg min H , c ∑ Δ ∈ L ( ln m ( Δ ) − η ) 2 − ln c − 2 H ln Δ ) 2 . (\widehat{H},\widehat{c})=\underset{H,c}{\operatorname{arg\,min}}\sum_{\Delta\in L}\left(\ln m(\Delta)-\eta\right)^2-\ln c-2H\ln\Delta)^2. ( H , c ) = H , c arg min Δ ∈ L ∑ ( ln m ( Δ ) − η ) 2 − ln c − 2 H ln Δ ) 2 .
Each series has its own noise constant,
η N = 2 N b − 1 , η R V = η ^ ( x r e f ‘ S e c t i o n 7.2 ‘ ) , η Q = 1 2 Var b [ ln Q b o d d − ln Q b e v e n ] , \eta_N=2N_b^{-1},\qquad\eta_{\mathrm{RV}}=\widehat{\eta}\quad({xref}`Section 7.2`),\qquad\eta_Q=\frac{1}{2}\operatorname{Var}_b[\ln Q_b^{\mathrm{odd}}-\ln Q_b^{\mathrm{even}}], η N = 2 N b − 1 , η RV = η ( x r e f ‘ S ec t i o n 7.2‘ ) , η Q = 2 1 Var b [ ln Q b odd − ln Q b even ] ,
and η = 0 \eta=0 η = 0 for the simulated series of § 6 . The constant η N \eta_N η N follows from Poisson counts with intensity λ b \lambda_b λ b in bin b b b ,
ln N b = ln λ b + N b − λ b λ b + O ( λ b − 1 ) , Var [ ln N b ∣ λ b ] = 1 λ b + O ( λ b − 2 ) , \ln N_b=\ln\lambda_b+\frac{N_b-\lambda_b}{\lambda_b}+O(\lambda_b^{-1}),\qquad\operatorname{Var}[\ln N_b\mid\lambda_b]=\frac{1}{\lambda_b}+O(\lambda_b^{-2}), ln N b = ln λ b + λ b N b − λ b + O ( λ b − 1 ) , Var [ ln N b ∣ λ b ] = λ b 1 + O ( λ b − 2 ) ,
E [ ( y b + Δ − y b ) 2 ] = m ln λ ( Δ ) + E [ 1 λ b + Δ + 1 λ b ] + O ( λ b − 2 ) ≈ m ln λ ( Δ ) + 2 N b − 1 . \mathbb{E}\left[(y_{b+\Delta}-y_b)^2\right]=m_{\ln\lambda}(\Delta)+\mathbb{E}\left[\frac{1}{\lambda_{b+\Delta}}+\frac{1}{\lambda_b}\right]+O(\lambda_b^{-2})\approx m_{\ln\lambda}(\Delta)+2N_b^{-1}. E [ ( y b + Δ − y b ) 2 ] = m l n λ ( Δ ) + E [ λ b + Δ 1 + λ b 1 ] + O ( λ b − 2 ) ≈ m l n λ ( Δ ) + 2 N b − 1 .
For Table 7 we estimate H ^ \widehat{H} H in each of the six blocks of 30 days and give the mean with the standard error over blocks.
Table A.1. Settings of the estimates of H H H .
Noise of the realised variance
Lemma A.1 (Noise of the realised variance). Let r b , i = σ b ε b , i r_{b,i}=\sigma_b\varepsilon_{b,i} r b , i = σ b ε b , i for i = 1 , … , n i=1,\ldots,n i = 1 , … , n , with ε b , i \varepsilon_{b,i} ε b , i i.i.d. N ( 0 , 1 ) N(0,1) N ( 0 , 1 ) over b b b and i i i and independent of ( σ b ) (\sigma_b) ( σ b ) , and let m Y ( Δ ) = E [ ( Y b + Δ − Y b ) 2 ] m_Y(\Delta)=\mathbb{E}\left[(Y_{b+\Delta}-Y_b)^2\right] m Y ( Δ ) = E [ ( Y b + Δ − Y b ) 2 ] . Then for every Δ ≥ 1 \Delta\geq1 Δ ≥ 1 ,
m ln R V ( Δ ) = m ln σ 2 ( Δ ) + 2 ψ 1 ( n 2 ) . m_{\ln RV}(\Delta)=m_{\ln\sigma^2}(\Delta)+2\psi_1\left(\frac{n}{2}\right). m l n R V ( Δ ) = m l n σ 2 ( Δ ) + 2 ψ 1 ( 2 n ) .
The noise of ln R V b \ln RV_b ln R V b thus adds the same constant at every lag. The idea of the proof is that ln R V b \ln RV_b ln R V b splits into ln σ b 2 \ln\sigma_b^2 ln σ b 2 and the logarithm of a Gamma variable that is independent across bins. The variance of that logarithm is the constant.
Proof of Lemma A.1. The noise as a Gamma variable. To separate the noise from σ b \sigma_b σ b , we factor σ b 2 \sigma_b^2 σ b 2 out of R V b RV_b R V b . As the ε b , i \varepsilon_{b,i} ε b , i are i.i.d. N ( 0 , 1 ) N(0,1) N ( 0 , 1 ) , their squares sum to a χ n 2 \chi_n^2 χ n 2 variable, which is twice a Γ ( n 2 , 1 ) \Gamma\left(\frac{n}{2},1\right) Γ ( 2 n , 1 ) variable G b G_b G b ,
R V b = σ b 2 ∑ i = 1 n ε b , i 2 = 2 σ b 2 G b , ln R V b = ln 2 + ln σ b 2 + ln G b . RV_b=\sigma_b^2\sum_{i=1}^{n}\varepsilon_{b,i}^2=2\sigma_b^2G_b,\qquad\ln RV_b=\ln2+\ln\sigma_b^2+\ln G_b. R V b = σ b 2 i = 1 ∑ n ε b , i 2 = 2 σ b 2 G b , ln R V b = ln 2 + ln σ b 2 + ln G b .
Different bins share no returns, so the G b G_b G b are independent over b b b . They are also independent of ( σ b ) (\sigma_b) ( σ b ) , as the ε b , i \varepsilon_{b,i} ε b , i are.
The variance of ln G b \ln G_b ln G b . The constant in the lemma is twice this variance, which we obtain from the cumulant generating function of ln G b \ln G_b ln G b . For t > − n 2 t>-\frac{n}{2} t > − 2 n , the Gamma density gives E [ G b t ] = Γ ( n 2 + t ) / Γ ( n 2 ) \mathbb{E}[G_b^t]=\Gamma\left(\frac{n}{2}+t\right)/\Gamma\left(\frac{n}{2}\right) E [ G b t ] = Γ ( 2 n + t ) /Γ ( 2 n ) . Taking logarithms and differentiating twice at t = 0 t=0 t = 0 , with ψ 1 = ( ln Γ ) ′ ′ \psi_1=(\ln\Gamma)'' ψ 1 = ( ln Γ ) ′′ , we obtain
K ( t ) = ln E [ e t ln G b ] = ln Γ ( n 2 + t ) − ln Γ ( n 2 ) , Var [ ln G b ] = K ′ ′ ( 0 ) = ψ 1 ( n 2 ) . K(t)=\ln\mathbb{E}\left[e^{t\ln G_b}\right]=\ln\Gamma\left(\frac{n}{2}+t\right)-\ln\Gamma\left(\frac{n}{2}\right),\qquad\operatorname{Var}[\ln G_b]=K''(0)=\psi_1\left(\frac{n}{2}\right). K ( t ) = ln E [ e t l n G b ] = ln Γ ( 2 n + t ) − ln Γ ( 2 n ) , Var [ ln G b ] = K ′′ ( 0 ) = ψ 1 ( 2 n ) .
The variogram. Let η b = ln G b − E [ ln G b ] \eta_b=\ln G_b-\mathbb{E}[\ln G_b] η b = ln G b − E [ ln G b ] . In the difference over a lag Δ ≥ 1 \Delta\geq1 Δ ≥ 1 , the constants ln 2 \ln2 ln 2 and E [ ln G b ] \mathbb{E}[\ln G_b] E [ ln G b ] cancel, which leaves
ln R V b + Δ − ln R V b = ( ln σ b + Δ 2 − ln σ b 2 ) + ( η b + Δ − η b ) . \ln RV_{b+\Delta}-\ln RV_b=\left(\ln\sigma_{b+\Delta}^2-\ln\sigma_b^2\right)+\left(\eta_{b+\Delta}-\eta_b\right). ln R V b + Δ − ln R V b = ( ln σ b + Δ 2 − ln σ b 2 ) + ( η b + Δ − η b ) .
We square and take expectations. The cross term vanishes, as ( η b ) (\eta_b) ( η b ) has mean zero and is independent of ( σ b ) (\sigma_b) ( σ b ) . The bins b b b and b + Δ b+\Delta b + Δ differ, so η b + Δ \eta_{b+\Delta} η b + Δ and η b \eta_b η b are independent, and each has variance ψ 1 ( n 2 ) \psi_1\left(\frac{n}{2}\right) ψ 1 ( 2 n ) . Hence
m ln R V ( Δ ) = m ln σ 2 ( Δ ) + 2 E [ ( ln σ b + Δ 2 − ln σ b 2 ) ( η b + Δ − η b ) ] + E [ ( η b + Δ − η b ) 2 ] = m ln σ 2 ( Δ ) + 0 + 2 ψ 1 ( n 2 ) . \begin{aligned}
m_{\ln RV}(\Delta)=m_{\ln\sigma^2}(\Delta)+2\mathbb{E}\left[(\ln\sigma_{b+\Delta}^2-\ln\sigma_b^2)(\eta_{b+\Delta}-\eta_b)\right]+\mathbb{E}\left[(\eta_{b+\Delta}-\eta_b)^2\right] \\
&=m_{\ln\sigma^2}(\Delta)+0+2\psi_1\left(\frac{n}{2}\right).
\end{aligned} m l n R V ( Δ ) = m l n σ 2 ( Δ ) + 2 E [ ( ln σ b + Δ 2 − ln σ b 2 ) ( η b + Δ − η b ) ] + E [ ( η b + Δ − η b ) 2 ] = m l n σ 2 ( Δ ) + 0 + 2 ψ 1 ( 2 n ) .
Returns with heavier tails than Gaussian raise Var [ ln ( R V b / σ b 2 ) ] \operatorname{Var}\left[\ln(RV_b/\sigma_b^2)\right] Var [ ln ( R V b / σ b 2 ) ] above ψ 1 ( n / 2 ) \psi_1(n/2) ψ 1 ( n /2 ) . The odd and the even minutes of a bin share σ b \sigma_b σ b , so the variance of their difference measures this rise,
r ^ = Var b [ ln R V b o d d − ln R V b e v e n ] ψ 1 ( ⌈ n / 2 ⌉ / 2 ) + ψ 1 ( ⌊ n / 2 ⌋ / 2 ) , η ^ = 2 ψ 1 ( n 2 ) r ^ . \hat{r}=\frac{\operatorname{Var}_b\left[\ln RV_b^{\mathrm{odd}}-\ln RV_b^{\mathrm{even}}\right]}{\psi_1(\lceil n/2\rceil/2)+\psi_1(\lfloor n/2\rfloor/2)}, \qquad\hat{\eta}=2\psi_1\left(\frac{n}{2}\right)\hat{r}. r ^ = ψ 1 (⌈ n /2 ⌉ /2 ) + ψ 1 (⌊ n /2 ⌋ /2 ) Var b [ ln R V b odd − ln R V b even ] , η ^ = 2 ψ 1 ( 2 n ) r ^ .
This step assumes that the tails raise the noise of n n n returns and of n / 2 n/2 n /2 returns by the same factor. A log-volatility with H = 0.105 H=0.105 H = 0.105 drives both series in a synthetic test (Table A.4). With Gaussian returns r ^ = 1.02 \hat{r}=1.02 r ^ = 1.02 , and the corrected H H H is 0.105. With Student t 3 t_3 t 3 returns r ^ = 2.61 \hat{r}=2.61 r ^ = 2.61 , and the corrected H H H is 0.065. On the six markets r ^ \hat{r} r ^ lies between 1.3 and 2.2.
Effect of the estimators on H
Three steps of our estimators change the fitted H H H : the removal of the pattern, the averaging within bins and the correction for noise. Removing the pattern from ln N b \ln N_b ln N b lowers H ( N b ) H(N_b) H ( N b ) by 0.03 to 0.05 (Table A.2). Terms of the daily pattern with periods shorter than 4 hours change it by at most 0.005, and the counting noise η N \eta_N η N by at most 0.001.
Table A.2. H ( N b ) H(N_b) H ( N b ) for the 180 days of Table 7, at bins of 5 minutes and lags of 15 minutes to one day, over the whole period.
To measure the effect of the bins, we give a log-intensity with known H H H one value per minute over 180 days and sum it in bins of 15, 30 and 60 minutes. Its fitted H H H rises with the bin length for every true H H H (Table A.3).
Table A.3. Fitted H H H of a log-intensity 2 X t 2X_t 2 X t with X X X a fractional Brownian motion of known H H H , at lags of 1 hour to 1 day.
For the correction η ^ \hat{\eta} η ^ of § 7.2 , one log-volatility drives both the counts and the returns of a synthetic market (Table A.4). With Gaussian returns η ^ \hat{\eta} η ^ restores the H H H of the counts. With Student t 3 t_3 t 3 returns it restores 0.065 of 0.105.
Table A.4. Synthetic market over 60 days of minutes. The log-variance ln ( σ t 2 / 10 − 6 ) \ln(\sigma_t^2/10^{-6}) ln ( σ t 2 /1 0 − 6 ) is a sum of 20 Ornstein–Uhlenbeck processes with time scales from 10 to 10 5 10^5 1 0 5 minutes, scaled to a standard deviation of 0.6. Counts are Poisson with mean 200 σ t 2 / 10 − 6 200\sigma_t^2/10^{-6} 200 σ t 2 /1 0 − 6 , and returns are σ t ε t \sigma_t\varepsilon_t σ t ε t . Bins of 15 minutes, lags of 1 hour to 1 day.
Kernel from the spectrum
Table 7 needs the norm a a a of the kernel, its lower bound and the tail α \alpha α . As the spectrum of a self-exciting process with mean rate λ ˉ \bar{\lambda} λ ˉ and kernel a g a\,g a g of norm ∥ g ∥ = 1 \lVert g\rVert=1 ∥ g ∥ = 1 is known in closed form [34 ] , we fit the kernel to the spectrum of the counts per second,
S ( ω ) = λ ˉ ∣ 1 − a g ^ ( ω ) ∣ 2 , S(\omega)=\frac{\bar{\lambda}}{\left|1-a\hat{g}(\omega)\right|^2}, S ( ω ) = ∣ 1 − a g ^ ( ω ) ∣ 2 λ ˉ ,
g ^ d e s k ( ω ) = ∑ k = 1 80 w k 1 + i ω T k , g ^ p o w ( ω ) = ∫ 0 ∞ α c α ( s + c ) − 1 − α e − i ω s d s . \hat{g}_{\mathrm{desk}}(\omega)=\sum_{k=1}^{80}\frac{w_k}{1+\mathrm{i}\omega T_k},\qquad\hat{g}_{\mathrm{pow}}(\omega)=\int_0^\infty\alpha c^\alpha(s+c)^{-1-\alpha}e^{-\mathrm{i}\omega s}\,\mathrm{d}s. g ^ desk ( ω ) = k = 1 ∑ 80 1 + i ω T k w k , g ^ pow ( ω ) = ∫ 0 ∞ α c α ( s + c ) − 1 − α e − i ω s d s .
Counts in bins of one second cannot tell the frequency ν \nu ν from ν + k \nu+k ν + k for integer k k k , and they weight each frequency by sinc 2 \operatorname{sinc}^2 sinc 2 . Relative to a Poisson process of the same rate, their spectrum is
f ( ν ) = w + A [ ∑ k = − 3 3 sinc 2 ( ν + k ) ( S ( 2 π ∣ ν + k ∣ ) λ ˉ − 1 ) + 1 ] , sinc x = sin π x π x , f(\nu)=w+A\left[\sum_{k=-3}^{3}\operatorname{sinc}^2(\nu+k)\left(\frac{S(2\pi|\nu+k|)}{\bar{\lambda}}-1\right)+1\right],\qquad\operatorname{sinc}x=\frac{\sin\pi x}{\pi x}, f ( ν ) = w + A [ k = − 3 ∑ 3 sinc 2 ( ν + k ) ( λ ˉ S ( 2 π ∣ ν + k ∣ ) − 1 ) + 1 ] , sinc x = π x sin π x ,
where w w w adds excitation within one second and A A A scales for the merging of fills (§ 7.1 ). As the rate of trading changes over a day, we compare f f f with the periodogram of segments s s s of 8 hours, each divided by its own mean x ˉ s \bar{x}_s x ˉ s ,
I ( ν ) = mean s ∣ ∑ t ( x s , t − x ˉ s ) e − 2 π i ν t ∣ 2 28800 x ˉ s . I(\nu)=\operatorname{mean}_{s}\frac{\left|\sum_t(x_{s,t}-\bar{x}_s)e^{-2\pi\mathrm{i}\nu t}\right|^2}{28800\bar{x}_s}. I ( ν ) = mean s 28800 x ˉ s ∑ t ( x s , t − x ˉ s ) e − 2 π i ν t 2 .
For frequencies ν \nu ν between 1 / ( 4 h ) 1/(4\,\mathrm{h}) 1/ ( 4 h ) and 1 / ( 1 m i n ) 1/(1\,\mathrm{min}) 1/ ( 1 min ) , the Whittle approximation of the negative log-likelihood is ∑ ν ( ln f ( ν ) + I ( ν ) / f ( ν ) ) \sum_\nu\left(\ln f(\nu)+I(\nu)/f(\nu)\right) ∑ ν ( ln f ( ν ) + I ( ν ) / f ( ν ) ) . To bound the norm from below, we minimise it over the other parameters at each fixed a a a ,
ℓ ( a ) = min α , c , A , w ∑ ν ( ln f ( ν ) + I ( ν ) f ( ν ) ) , \ell(a)=\min_{\alpha,c,A,w}\sum_\nu\left(\ln f(\nu)+\frac{I(\nu)}{f(\nu)}\right), ℓ ( a ) = α , c , A , w min ν ∑ ( ln f ( ν ) + f ( ν ) I ( ν ) ) ,
a ^ = arg min a ℓ ( a ) , a 95 = min { a : ℓ ( a ) − ℓ ( a ^ ) ≤ 1.92 } . \hat{a}=\arg\min_a\ell(a),\qquad a_{95}=\min\left\{a:\ell(a)-\ell(\hat{a})\le1.92\right\}. a ^ = arg a min ℓ ( a ) , a 95 = min { a : ℓ ( a ) − ℓ ( a ^ ) ≤ 1.92 } .
with 1.92 half the 95% quantile of χ 1 2 \chi^2_1 χ 1 2 . The gain in Table 7 is ℓ p o w ( a ^ ) − ℓ d e s k ( a ^ ) \ell_{\mathrm{pow}}(\hat{a})-\ell_{\mathrm{desk}}(\hat{a}) ℓ pow ( a ^ ) − ℓ desk ( a ^ ) . For the standard error of α \alpha α we repeat the fit on each third of the period.
Table A.5. Settings of the fit of the kernel.
Experiments
Settings of the simulations
Table B.1. Settings shared by all simulations.
Table B.2. One experiment per column.
In every experiment we obtain β \beta β by dividing β ˉ κ \bar{\beta}\kappa β ˉ κ in Table B.2 by the measured κ ˉ \bar{\kappa} κ ˉ .
Assumption 3.3 and the constant kappa along the runs
The linear response of § 6.3 and the limits of § 5 rest on Assumption 3.3 and on a constant κ ˉ \bar{\kappa} κ ˉ (§ 4.1 ). In the notation of § 2 , these three conditions read
max k p k ( n ) < 1 , min i , j ( x i , j ( n ) − σ i θ n ) > 0 , κ l ( n ) = κ l . \max_k p_k(n) < 1,\qquad\min_{i,j}\left(x_{i,j}(n)-\sigma_i\theta_n\right)>0,\qquad\kappa_l(n)=\kappa_l. k max p k ( n ) < 1 , i , j min ( x i , j ( n ) − σ i θ n ) > 0 , κ l ( n ) = κ l .
We measure each condition along every run of § 6.2 to § 6.5 .
Table B.3. The conditions of Assumption 3.3 and the constancy of κ ˉ \bar{\kappa} κ ˉ (2 ⋅ 10 6 2\cdot10^6 2 ⋅ 1 0 6 steps per run; the last column gives the range over the eight values of s s s ).
The limit of a single cluster
Lemma B.1 (Roughness in the limit of a single cluster). Let X X X solve d X t = a ( μ − X t ) d t + σ X t d B t dX_t=a(\mu-X_t)\,dt+\sigma\sqrt{X_t}\,dB_t d X t = a ( μ − X t ) d t + σ X t d B t with a , μ , σ > 0 a,\mu,\sigma>0 a , μ , σ > 0 , started from its stationary law, and let k = 2 a μ / σ 2 > 2 k=2a\mu/\sigma^2>2 k = 2 a μ / σ 2 > 2 . Then
m ( Δ ) = E [ ( ln X t + Δ − ln X t ) 2 ] = 2 a k − 1 Δ + O ( Δ 3 / 2 ) , Δ → 0 , m(\Delta)=\mathbb{E}\left[(\ln X_{t+\Delta}-\ln X_t)^2\right]=\frac{2a}{k-1}\Delta+O(\Delta^{3/2}),\qquad\Delta\to0, m ( Δ ) = E [ ( ln X t + Δ − ln X t ) 2 ] = k − 1 2 a Δ + O ( Δ 3/2 ) , Δ → 0 , and hence H = 1 2 H=\frac{1}{2} H = 2 1 .
The shape of the stationary law is k = ( E X ) 2 / Var X k=(\mathbb{E}X)^2/\operatorname{Var}X k = ( E X ) 2 / Var X . In the run with a single cluster of Table B.2 we measure k = 2.19 k=2.19 k = 2.19 (s.e. 0.09).
Proof of Lemma B.1.
X t ∼ Γ ( k , 2 a σ 2 ) , E [ X t − 1 ] = 2 a σ 2 ( k − 1 ) , E [ X t − 2 ] < ∞ , X_t\sim\Gamma\left(k,\frac{2a}{\sigma^2}\right),\qquad\mathbb{E}[X_t^{-1}]=\frac{2a}{\sigma^2(k-1)},\qquad\mathbb{E}[X_t^{-2}]<\infty, X t ∼ Γ ( k , σ 2 2 a ) , E [ X t − 1 ] = σ 2 ( k − 1 ) 2 a , E [ X t − 2 ] < ∞ ,
d ln X t = g ( X t ) d t + σ X t − 1 / 2 d B t , g ( x ) = a μ − σ 2 / 2 x − a , d\ln X_t=g(X_t)\,dt+\sigma X_t^{-1/2}\,dB_t,\qquad g(x)=\frac{a\mu-\sigma^2/2}{x}-a, d ln X t = g ( X t ) d t + σ X t − 1/2 d B t , g ( x ) = x a μ − σ 2 /2 − a ,
ln X t + Δ − ln X t = ∫ t t + Δ g ( X s ) d s ⏟ D Δ + σ ∫ t t + Δ X s − 1 / 2 d B s ⏟ M Δ , \ln X_{t+\Delta}-\ln X_t=\underbrace{\int_t^{t+\Delta}g(X_s)\,ds}_{D_\Delta}+\sigma\underbrace{\int_t^{t+\Delta}X_s^{-1/2}\,dB_s}_{M_\Delta}, ln X t + Δ − ln X t = D Δ ∫ t t + Δ g ( X s ) d s + σ M Δ ∫ t t + Δ X s − 1/2 d B s ,
E [ M Δ 2 ] = σ 2 ∫ t t + Δ E [ X s − 1 ] d s = 2 a k − 1 Δ , \mathbb{E}[M_\Delta^2]=\sigma^2\int_t^{t+\Delta}\mathbb{E}[X_s^{-1}]\,ds=\frac{2a}{k-1}\Delta, E [ M Δ 2 ] = σ 2 ∫ t t + Δ E [ X s − 1 ] d s = k − 1 2 a Δ ,
E [ D Δ 2 ] ≤ Δ ∫ t t + Δ E [ g ( X s ) 2 ] d s = E [ g ( X t ) 2 ] Δ 2 , \mathbb{E}[D_\Delta^2]\leq\Delta\int_t^{t+\Delta}\mathbb{E}[g(X_s)^2]\,ds=\mathbb{E}[g(X_t)^2]\Delta^2, E [ D Δ 2 ] ≤ Δ ∫ t t + Δ E [ g ( X s ) 2 ] d s = E [ g ( X t ) 2 ] Δ 2 ,
∣ E [ D Δ M Δ ] ∣ ≤ E [ D Δ 2 ] 1 / 2 E [ M Δ 2 ] 1 / 2 = O ( Δ 3 / 2 ) , \left|\mathbb{E}[D_\Delta M_\Delta]\right|\leq\mathbb{E}[D_\Delta^2]^{1/2}\mathbb{E}[M_\Delta^2]^{1/2}=O(\Delta^{3/2}), ∣ E [ D Δ M Δ ] ∣ ≤ E [ D Δ 2 ] 1/2 E [ M Δ 2 ] 1/2 = O ( Δ 3/2 ) ,
m ( Δ ) = E [ M Δ 2 ] + 2 E [ D Δ M Δ ] + E [ D Δ 2 ] = 2 a k − 1 Δ + O ( Δ 3 / 2 ) . m(\Delta)=\mathbb{E}[M_\Delta^2]+2\mathbb{E}[D_\Delta M_\Delta]+\mathbb{E}[D_\Delta^2]=\frac{2a}{k-1}\Delta+O(\Delta^{3/2}). m ( Δ ) = E [ M Δ 2 ] + 2 E [ D Δ M Δ ] + E [ D Δ 2 ] = k − 1 2 a Δ + O ( Δ 3/2 ) .
Data of Binance
Table B.4. Trades against resting volume on the perpetual futures, 1–3 October 2026.
Algorithm B.1 Change of the book caused by the move of the mid price
Require: snapshots n n n at times t n t_n t n with cumulative volumes C n ± ( x k ) C_n^{\pm}(x_k) C n ± ( x k ) on the ask (+) and bid (−) sides, x k ∈ { 0 , 0.2 , 1 , 2 , 3 , 4 , 5 } % x_k \in\{0, 0.2, 1, 2, 3, 4, 5\}\% x k ∈ { 0 , 0.2 , 1 , 2 , 3 , 4 , 5 } % , C n ± ( 0 ) = 0 C_n^{\pm}(0)=0 C n ± ( 0 ) = 0 ; last trade prices S t S_t S t
Ensure: Δ V n ( d ) \Delta V_n(d) Δ V n ( d ) and Δ n m i d ( d ) \Delta_n^{\mathrm{mid}}(d) Δ n mid ( d ) for d = 1 , … , 4 d=1,\ldots,4 d = 1 , … , 4 on both sides; slope ς \varsigma ς
for each pair of snapshots ( n , n + 1 ) (n,n+1) ( n , n + 1 ) do
skip unless 0 < t n + 1 − t n ≤ 120 s 0 < t_{n+1}-t_n \le120\,\mathrm{s} 0 < t n + 1 − t n ≤ 120 s , C n ± C_n^{\pm} C n ± and C n + 1 ± C_{n+1}^{\pm} C n + 1 ± increase in x x x with C ± ( 0.2 ) > 0 C^{\pm}(0.2)>0 C ± ( 0.2 ) > 0 , C n + 1 ≠ C n C_{n+1}\ne C_n C n + 1 = C n , and S t n , S t n + 1 S_{t_n},S_{t_{n+1}} S t n , S t n + 1 exist
μ n ← 100 ( ln S t n + 1 − ln S t n ) \mu_n \leftarrow100(\ln S_{t_{n+1}}-\ln S_{t_n}) μ n ← 100 ( ln S t n + 1 − ln S t n ) ▷ move of the mid price in %
C ^ n ± ( x ) ← C n ± ( x k ) + x − x k x k + 1 − x k ( C n ± ( x k + 1 ) − C n ± ( x k ) ) \widehat{C}_n^{\pm}(x)\leftarrow C_n^{\pm}(x_k)+\frac{x-x_k}{x_{k+1}-x_k}\left(C_n^{\pm}(x_{k+1})-C_n^{\pm}(x_k)\right) C n ± ( x ) ← C n ± ( x k ) + x k + 1 − x k x − x k ( C n ± ( x k + 1 ) − C n ± ( x k ) ) , k = min { max { k : x k ≤ x } , 5 } k=\min\{\max\{k:x_k\le x\},5\} k = min { max { k : x k ≤ x } , 5 }
C n + , μ ( x ) ← C ^ n + ( x + μ n + ) − C ^ n + ( μ n + ) C_n^{+,\mu}(x)\leftarrow\widehat{C}_n^{+}(x+\mu_n^{+})-\widehat{C}_n^{+}(\mu_n^{+}) C n + , μ ( x ) ← C n + ( x + μ n + ) − C n + ( μ n + ) ▷ orders keep their price, crossed orders leave
C n − , μ ( x ) ← C ^ n − ( x − μ n − ) − C ^ n − ( ( − μ n ) + ) C_n^{-,\mu}(x)\leftarrow\widehat{C}_n^{-}(x-\mu_n^{-})-\widehat{C}_n^{-}((-\mu_n)^{+}) C n − , μ ( x ) ← C n − ( x − μ n − ) − C n − (( − μ n ) + )
Δ n m i d ( d ) ← [ C n m i d ( d + 1 ) − C n m i d ( d ) ] − [ C n ( d + 1 ) − C n ( d ) ] \Delta_n^{\mathrm{mid}}(d)\leftarrow[C_n^{\mathrm{mid}}(d+1)-C_n^{\mathrm{mid}}(d)]-[C_n(d+1)-C_n(d)] Δ n mid ( d ) ← [ C n mid ( d + 1 ) − C n mid ( d )] − [ C n ( d + 1 ) − C n ( d )] ▷ each side
Δ V n ( d ) ← [ C n + 1 ( d + 1 ) − C n + 1 ( d ) ] − [ C n ( d + 1 ) − C n ( d ) ] \Delta V_n(d)\leftarrow[C_{n+1}(d+1)-C_{n+1}(d)]-[C_n(d+1)-C_n(d)] Δ V n ( d ) ← [ C n + 1 ( d + 1 ) − C n + 1 ( d )] − [ C n ( d + 1 ) − C n ( d )]
end for
ς ← ∑ n , d , ± Δ V n ( d ) Δ n m i d ( d ) ∑ n , d , ± Δ n m i d ( d ) 2 \varsigma\leftarrow\frac{\sum_{n,d,\pm}\Delta V_n(d)\,\Delta_n^{\mathrm{mid}}(d)}{\sum_{n,d,\pm}\Delta_n^{\mathrm{mid}}(d)^2} ς ← ∑ n , d , ± Δ n mid ( d ) 2 ∑ n , d , ± Δ V n ( d ) Δ n mid ( d ) ▷ kept pairs, d = 1 , … , 4 d=1,\ldots,4 d = 1 , … , 4 , both sides
Desks compared with Binance
To compare the six markets with our desks, we simulate a self-exciting process with their kernel for § 7.2 and their books for § 7.3 , with the settings of Table B.5.
Table B.5. Settings of the desks compared with Binance.
Table B.6. H ( N b ) H(N_b) H ( N b ) of the self-exciting process with the kernel of the desks.
Table B.7. ρ \rho ρ and ς \varsigma ς of the books of the desks.