A fully-discrete Semi-Lagrangian scheme for a first order mean field game problem

• The first-order system couples a backward HJB equation with terminal data to a forward continuity equation transporting the initial measure by the flow induced by −Dv.The value function is expected to be only Lipschitz, while the density is expected to have L∞ regularity.Source: §1 · p. 2
• For coupled equations, the authors identify only forward-forward numerical schemes in prior work, whose initial-value structure changes the analysis from this forward-backward problem.Source: §1 · p. 2
• The proposed fully discrete Semi-Lagrangian scheme admits at least one solution, and every limit point converges to a solution of (1.2) when d = 1 and discretization parameters tend to zero suitably.The scheme is formulated in general dimension, and several of its properties are proved in that generality.Source: §1 · p. 3
• The convergence proof relies on discrete semiconcavity and uniform L∞ bounds, which the authors establish only for d = 1.Condition (1.3) corresponds to semiconvexity of v and does not hold for arbitrary time horizons.Source: §1 · pp. 2–3

2.1 Basic assumptions and existence and uniqueness results for (1.2)

• Under (H1)–(H3), system (1.2) has a solution; adding monotonicity conditions (H4) for F and G yields uniqueness.The assumptions include continuous F and G and an absolutely continuous, bounded initial density with bounded support.Source: §2.1 · p. 4

2.2 Standard semiconcavity results

• Semiconcavity supplies key regularity: such functions are locally Lipschitz, and sequences with a shared semiconcavity constant converge locally uniformly with gradients converging almost everywhere.The paper also characterizes semiconcavity through the distributional bound D2w ≤ CconcId.Source: §2.2 · pp. 4–5

2.3 Representation formulas for the solutions of the HJB and the continuity equations

• For prescribed µ, selected optimal trajectories follow the flow equation driven by −Dv[µ], and transporting m0 along that flow yields the unique distributional solution of the continuity equation.The transported measure remains absolutely continuous, supported in a bounded ball, and uniformly bounded in L∞.Source: §2.3 · p. 7

3 The fully-discrete scheme

• The spatial discretization uses compactly supported nodal P1 basis functions whose sum is one.Each basis function equals one at its own lattice node and zero at the others.Source: §3 · p. 8

3.1 The fully-discrete scheme for the HJB equation

• The HJB update is well defined, uniformly bounded, and monotone in its value-function input.The scheme also yields uniform-in-grid Lipschitz and weak semiconcavity properties.Source: §3.1 · p. 9
• The text states that, after establishing the subsolution and supersolution properties, a classical comparison argument yields local uniform convergence of vρn,hn[µn] to v[µ] in Rd × (0,T).Source: §3.1 · p. 11

3.2 The fully-discrete scheme for the continuity equation

• The continuity-equation discretization is conservative; when ρ = O(h), its evolving measures have uniformly bounded support.The support bound is independent of the discretization parameters and µ.Source: §3.2 · pp. 13, 15

3.3 The fully-discrete scheme for the first order MFG problem (1.2)

• The section defines a full discretization of the first-order MFG system using the discrete HJB and continuity-equation components.The formulation is given in equation (3.31).Source: §3.3 · p. 17
• The fully discrete system (3.31) admits at least one solution, by the stability lemma, Proposition 3.2(ii), and Brouwer’s fixed-point theorem.Source: §3.3 · p. 17
• In d = 1, every limit point of discrete solutions solves (MFG) when ρ_n = o(h_n), h_n = o(ε_n), and ρ_n = O(ε_n^1.5) as ε_n ↓ 0.The theorem assumes (H1)–(H3).Source: §3.3 · p. 18
• Under (H4), the discrete solutions converge in C([0, T]; P1) to the unique solution of (MFG).Source: §3.3 · p. 18

4 Numerical Tests

• The paper does not analyze whether the fixed-point iterations converge; developing a convergent and efficient solver remains future work.The tests use fixed-point iteration to compute the mass and value functions.Source: §4 · p. 20

4.1 Test 1

• With 20 iterations per parameter set, errors trend upward as ε decreases because finer discretizations require more iterations to reach the fixed error threshold.The parameter sets follow Theorem 3.4’s balancing requirements.Source: §4.1 · p. 23
• In Test 1, the mass-distribution errors decrease as the number of fixed-point iterations increases.Figure 2 (left) shows errors while all parameters vary according to Table 1.Source: Figure 2 · p. 22, §4.1 · p. 23
• Although Theorem 3.4 requires ε to be large relative to ρ, the tests still show numerical convergence with smaller ε.Figure 2 (right) varies ε at the (ρ, h) settings of Table 1.Source: Figure 2 · p. 22, §4.1 · p. 23

4.2 Test 2

• During mass evolution, the distribution tends to concentrate at the “low energy” configuration x = 0.2, while the second term in F penalizes high mass concentrations.Source: §4.2 · pp. 24–25
• Without a game coupling, the system reaches its solution after one iteration and the mass can concentrate, unlike the coupled case.The paper attributes this qualitative difference to the absence of a high-mass penalization term.Source: Figure 5 · p. 24, §4.2 · p. 25