Exact Solvability of the Correlated Price-Default Portfolio

This page derives, from first principles, what can be solved exactly for the correlated price-default portfolio problem stated in the research plan. The plan poses three research lemmas (L1 constant intensity, L2 affine price-dependent intensity, L3 general correlated intensity) and asks for the solvability boundary to be stated as a theorem rather than left implicit.

The decisive fact is that the solvability boundary depends on the utility function, not on the intensity specification alone. The default is a multiplicative wealth jump, \(X \mapsto X(1-\pi_i(1-R_i))\), and each utility family transforms that jump into a different algebraic object. The page states the general problem once, derives the reduction under three utility families (CARA, log/Kelly, power/CRRA), and closes with a matrix locating each lemma in each family.

Throughout, notation follows the plan and Exact Solutions in Stochastic Optimal Control; \(a(t) = e^{r(T-t)}\) is the risk-free discount factor.

Notation and assumptions

SymbolMeaning
\(m\)number of risky assets
\(S_i\)price of asset \(i\), correlated geometric Brownian motion
\(D_i \in \lbrace0,1\rbrace\)default indicator; \(D_i = 0\) alive, \(D_i = 1\) defaulted
\(X\)wealth; risk-free asset pays rate \(r \ge 0\)
\(\pi_i\)fraction of wealth in asset \(i\)
\(\phi_i = \pi_i X\)dollar amount in asset \(i\) (the natural CARA control, introduced below)
\(\mu_i, \sigma_i\)constant drift and volatility of asset \(i\)
\(\Sigma_{ij} = \rho_{ij}\sigma_i\sigma_j\)constant covariance matrix
\(\lambda_i(S,D)\)default intensity of asset \(i\)
\(R_i \in [0,1)\)recovery rate on default of asset \(i\)
\(\gamma > 0\)CARA risk-aversion coefficient
\(\eta > 0,\ \eta \ne 1\)CRRA relative risk-aversion coefficient
\(a(t) = e^{r(T-t)}\)risk-free discount factor, \(a(T) = 1\)
\(U(x)\)generic increasing concave utility, in the general statement

Model assumptions. Constant covariance \(\Sigma\), constant \(\mu_i\), \(\sigma_i\), constant recovery \(R_i\), finite horizon \(T\).

Technical assumptions. Bounded intensities, finite default-state space \(D \in \lbrace0,1\rbrace^m\), integrable terminal condition. These do not affect the algebraic structure.

The general problem

Alive assets follow correlated geometric Brownian motion,

\[ dS_i = S_i\left(\mu_i\,dt + \sigma_i\,dW_i\right), \qquad d\langle W_i, W_j\rangle_t = \rho_{ij}\,dt, \qquad \Sigma_{ij} = \rho_{ij}\sigma_i\sigma_j . \]

Wealth evolves as

\[ dX = X\Big(r + \sum_i \pi_i(\mu_i - r)\Big)dt + X\sum_i \pi_i\sigma_i\,dW_i , \]

where the sums run over alive assets (\(D_i = 0\)) only, and on default of asset \(i\),

\[ X \mapsto X\big(1 - \pi_i(1-R_i)\big), \qquad D \mapsto D + e_i . \]

The loss \(X\pi_i(1-R_i)\) is multiplicative in wealth: it scales the whole position by \((1-\pi_i(1-R_i))\). This single fact drives every result.

The objective is a generic increasing concave utility of terminal wealth,

\[ V(t,S,D,X) = \sup_\pi \mathbb{E}\!\left[\,U(X_T)\,\right], \]

with terminal condition \(V(T,S,D,X) = U(X)\). The HJB equation is

\[ 0 = \partial_t V + \sup_\pi\Bigg[ X\Big(r + \sum_i \pi_i(\mu_i - r)\Big)\partial_X V

  • \tfrac12 X^2\sum_{i,j}\pi_i\pi_j\Sigma_{ij}\,\partial_{XX}V
  • \sum_i \lambda_i(S,D)\Big(V\big(t,S,D+e_i,\,X(1-\pi_i(1-R_i))\big) - V\Big) \Bigg]. \]

Only the jump term depends on the utility, and only through the shifted value \(V(t,S,D+e_i,\,X(1-\pi_i(1-R_i)))\). The table below shows how each family factors that term under its separation ansatz; this is the structural fork that decides solvability.

Utility\(U(x)\)Value ansatzwealth-independent controldefault term in the FOC
CARA\(-e^{-\gamma x}\)\(-e^{-\gamma X a}\,u(t,S,D)\)dollar amount \(\phi_i = \pi_i X\)\(e^{\gamma a\,\phi_i(1-R_i)}\) (transcendental)
Log (Kelly)\(\log x\)\(\log X + g(t,S,D)\)fraction \(\pi_i\)\(\dfrac{1-R_i}{1-\pi_i(1-R_i)}\) (rational)
Power (CRRA)\(\dfrac{x^{1-\eta}}{1-\eta}\)\(\dfrac{X^{1-\eta}}{1-\eta}\,h(t,S,D)\)fraction \(\pi_i\)\((1-\pi_i(1-R_i))^{-\eta}\) (algebraic)

CARA (exponential) utility

Take \(U(x) = -e^{-\gamma x}\) and the ansatz

\[ V(t,S,D,X) = -\exp\!\big(-\gamma X a(t)\big)\; u(t,S,D), \qquad u(T,S,D) = 1 . \]

The wealth derivatives are

\[ \partial_X V = \gamma a\, u\, e^{-\gamma X a}, \qquad \partial_{XX}V = -\gamma^2 a^2\, u\, e^{-\gamma X a}, \qquad \partial_t V = -e^{-\gamma X a}\big[\partial_t u + \gamma X r a\, u\big], \]

and the jump term becomes

\[ V\big(t,S,D+e_i,\,X(1-\pi_i(1-R_i))\big) - V = e^{-\gamma X a}\Big[u - e^{\gamma X a \pi_i(1-R_i)}\,u\big(t,S,D+e_i\big)\Big]. \]

Two observations follow. First, the factor \(\exp(\gamma X a \pi_i(1-R_i))\) still contains \(X\), so the jump does not factor under the fraction control \(\pi_i\); the plan's claim that "the first-order condition for \(\pi\) is homogeneous in \(X\), so \(X\) cancels" is incorrect for a multiplicative jump. Second, \(X\) disappears iff the control is the dollar amount \(\phi_i = \pi_i X\), since then

\[ X\pi_i(\mu_i - r) \to \phi_i(\mu_i - r), \qquad X^2\pi_i\pi_j \to \phi_i\phi_j, \qquad e^{\gamma X a\pi_i(1-R_i)} \to e^{\gamma a\,\phi_i(1-R_i)} . \]

The correct CARA separation is therefore in dollar amounts: \(\phi_i\) is wealth-independent and the optimal fraction is \(\pi_i^* = \phi_i^*/X\).

The reduced equation is

\[ \begin{aligned} 0 ={}& -\partial_t u - \sum_{i\in\mathrm{Alive}(D)} \mu_i S_i\,\partial_{S_i}u - \tfrac12 \sum_{i,j\in\mathrm{Alive}(D)} \Sigma_{ij} S_i S_j\,\partial_{S_i S_j}u \\ & + \gamma a \sum_{k\in\mathrm{Alive}(D)} \phi_k(\mu_k - r)\,u - \tfrac12\gamma^2 a^2 \sum_{k,l\in\mathrm{Alive}(D)} \phi_k\phi_l\Sigma_{kl}\,u \\ & + \gamma a \sum_{i,j\in\mathrm{Alive}(D)} \phi_j \Sigma_{ij} S_i\,\partial_{S_i}u + \\ & \sum_i \lambda_i(S,D)\Big[u - e^{\gamma a\,\phi_i(1-R_i)}\,u(t,S,D+e_i)\Big], \end{aligned} \]

and the first-order condition in \(\phi_j\) is

\[ (\mu_j - r) - \gamma a\sum_k \Sigma_{jk}\,\phi_k

  • \frac{1}{u}\sum_i \Sigma_{ij} S_i\,\partial_{S_i}u \;=\; \lambda_j(S,D)\,(1-R_j)\,e^{\gamma a\,\phi_j(1-R_j)}\;\frac{u(t,S,D+e_j)}{u(t,S,D)} . \]

The first two \(S\)-terms are the price generator carried by the state \(S\); the term with \(S_i\partial_{S_i}u\) in the first-order condition is the price-wealth cross term. Both vanish in L1 because the value is independent of \(S\), and both must be carried in L2/L3.

With no default (\(\lambda_i=0\)) this reduces to the Merton dollar position \(\phi^* = \tfrac{1}{\gamma a}\Sigma^{-1}(\mu-r)\) and \(u = \exp(-\tfrac12(\mu-r)^\top\Sigma^{-1}(\mu-r)(T-t))\), the classical CARA Merton value [@merton1969lifetime]. The default term \(e^{\gamma a\,\phi_j(1-R_j)}\) is transcendental in \(\phi_j\) and is coupled to the adjacent-state ratio \(u(t,S,D+e_j)/u(t,S,D)\); it is the source of the nonlinearity and it survives in every intensity regime, including L1.

L1 numerical reduction

The L1 system is triangular in the default coordinate: the equation for a state \(D\) depends on \(u(t,D+e_i)\), a state with one more default. Ordering the states by decreasing number of defaults turns the coupled system into one scalar backward ODE per state. In remaining horizon \(\tau = T-t\),

\[ \partial_\tau u = -A(D,u)\,u - \sum_i \lambda_i \big(u - e^{\gamma a\phi_i(1-R_i)}u(\tau,D+e_i)\big), \]

with

\[ A(D,u) = \gamma a \sum_{i \text{ alive}} \Big(\phi_i(\mu_i-r) - \tfrac12\gamma a\sigma_i^2\phi_i^2\Big), \]

and \(\phi_i\) the root of the first-order condition above. The implementation orders the states and integrates each scalar ODE with RK4; the root is found by bisection.

Two exact limits validate the solver:

  • No default \((\lambda_i = 0)\): \(u(\tau,D) = \exp\big(-\tfrac12\tau \sum_{i \text{ alive}} (\mu_i-r)^2/\sigma_i^2\big)\), the Merton CARA value.
  • No loss \((R_i = 1)\): the first-order condition decouples, \(\phi_i\) is the Merton dollar position, and the reduced problem is linear: \[ u(\tau) = \exp\big((\Lambda - \operatorname{diag}(A))\tau\big)\mathbf{1}. \]

Both limits are tested against the nonlinear solver; the L1 case itself is checked for convergence under step refinement. The implementation is neural_solver/models/jump_to_default_cara.py.

Log utility (Kelly criterion)

Take \(U(x) = \log x\) and the additive ansatz

\[ V(t,S,D,X) = \log X + g(t,S,D), \qquad g(T,S,D) = 0 . \]

Log utility is scale-invariant, so the fraction \(\pi_i\) (not a dollar amount) is the wealth-independent control. Substituting \(\partial_X V = 1/X\) and \(\partial_{XX}V = -1/X^2\), the multiplicative jump becomes additive:

\[ \log\!\big(X(1-\pi_i(1-R_i))\big) + g(t,S,D+e_i) - \log X - g(t,S,D) = \log(1-\pi_i(1-R_i)) + g(t,S,D+e_i) - g(t,S,D). \]

The reduced equation is

\[ 0 = \partial_t g + \sum_i \lambda_i(S,D)\big(g(t,S,D+e_i)-g(t,S,D)\big) + r + \sup_\pi\Big[ \sum_i\pi_i(\mu_i-r) - \tfrac12\sum_{i,j}\pi_i\pi_j\Sigma_{ij}

  • \sum_i\lambda_i(S,D)\log(1-\pi_i(1-R_i))\Big], \]

and the crucial point is that the value-difference \(g(t,S,D+e_i)-g(t,S,D)\) is independent of \(\pi\), so the supremum is a standalone growth-rate maximization decoupled from \(g\). Its first-order condition is

\[ (\mu_j - r) - \sum_k \Sigma_{jk}\,\pi_k \;=\; \lambda_j(S,D)\,\frac{1-R_j}{1-\pi_j(1-R_j)} . \]

This is rational in \(\pi_j\) (quadratic for a single asset, the "Merton with deterministic jumps" case of Exact Solutions in Stochastic Optimal Control), and its solution \(\pi^*(S,D)\) depends on the state only through \(\lambda_i(S,D)\) — not on \(g\), \(t\), or \(X\). Denoting the maximized constant \(G(t,S,D) = r + \sup_\pi[\cdots]\), the value \(g\) solves the linear equation

\[ \partial_t g(t,S,D) + \sum_i \lambda_i(S,D)\big(g(t,S,D+e_i)-g(t,S,D)\big) = -G(t,S,D), \qquad g(T,S,D) = 0 . \]

For constant intensities (L1) both \(\pi^*\) and \(G\) are constants and this is a linear, constant-coefficient nonhomogeneous system over the \(2^m\) default states, solvable in closed form by the matrix exponential of the default-chain generator \(\Lambda\),

\[ g(t,D) = \int_0^{T-t} \big(e^{\Lambda s} G\big)(D)\,ds, \qquad (\Lambda g)(D) = \sum_i \lambda_i\big(g(D+e_i) - g(D)\big), \]

where \(G = G(D)\) is the maximized growth over the alive assets. If \(G\) is the same at every state, this reduces to \(g(t,D) = G\int_0^{T-t} e^{\Lambda s} \mathbf{1}\,ds\). The state-dependent case is the one implemented in neural_solver.models.jump_to_default, by the augmented matrix exponential

\[ \exp\!\left(\begin{bmatrix}\Lambda & G\\ 0 & 0\end{bmatrix}\tau\right) = \begin{bmatrix} e^{\Lambda\tau} & \int_0^{\tau} e^{\Lambda s} G\,ds\\ 0 & 1\end{bmatrix}. \]

For affine or general intensities (L2, L3) the equation remains linear in \(g\), with state-dependent coefficients, and is solvable by Feynman–Kac. This is the exact matrix-exponential structure the plan sought for L1 — but it holds under log utility, not CARA.

The default generator on the Boolean cube

The default state is \(D \in \mathbb{Z}_2^m\). The generator of a single default is translation by \(e_i\) minus the identity,

\[ \mathcal{J}_i u(D) = u(D+e_i) - u(D). \]

With the Walsh character \(\chi_\omega(D) = (-1)^{\omega \cdot D}\), and \(\chi_\omega(D+e_i) = (-1)^{\omega_i}\chi_\omega(D)\), the Walsh-Hadamard transform diagonalizes it:

\[ \widehat{\mathcal{J}_i u}(\omega) = \big((-1)^{\omega_i} - 1\big)\,\hat u(\omega). \]

The \(D\)-independent part of the default generator is therefore a Walsh multiplier with symbol \(\sum_i \lambda_i((-1)^{\omega_i}-1)\). A contagion intensity \(\lambda_i(S,D)\) is a multiplication operator, which becomes a convolution on the character group, so the generator is a full or low-rank matrix over characters rather than a diagonal symbol. The absorbing condition, a defaulted asset cannot default again, is a pointwise mask on the dead states and is the sub-Markovian killing term discussed in the validity note. This is the structure the Walsh operator encodes; the alive-masked generator is what the exact L1 lattice uses.

Power (CRRA) utility

Take \(U(x) = x^{1-\eta}/(1-\eta)\), \(\eta > 0\), \(\eta \ne 1\), and the homogeneous ansatz

\[ V(t,S,D,X) = \frac{X^{1-\eta}}{1-\eta}\,h(t,S,D), \qquad h(T,S,D) = 1 . \]

Homogeneity makes the fraction \(\pi_i\) wealth-independent, as in the log case. With \(\partial_X V = X^{-\eta}h\) and \(\partial_{XX}V = -\eta X^{-\eta-1}h\), the jump term becomes a power,

\[ V\big(t,S,D+e_i,\,X(1-\pi_i(1-R_i))\big) - V = \frac{X^{1-\eta}}{1-\eta}\Big[(1-\pi_i(1-R_i))^{1-\eta}h(t,S,D+e_i) - h\Big], \]

and, multiplying through by \((1-\eta)/X^{1-\eta}\), the reduced equation is

\[ 0 = \partial_t h + \sup_\pi\Big[ (1-\eta)\Big(r+\sum_i\pi_i(\mu_i-r)\Big)h

  • \tfrac{\eta(1-\eta)}{2}\sum_{i,j}\pi_i\pi_j\Sigma_{ij}\,h
  • \sum_i \lambda_i\Big((1-\pi_i(1-R_i))^{1-\eta}h(t,S,D+e_i) - h\Big)\Big]. \]

The first-order condition in \(\pi_j\) is

\[ (\mu_j - r) - \eta\sum_k \Sigma_{jk}\,\pi_k \;=\; \lambda_j(S,D)\,(1-R_j)\,(1-\pi_j(1-R_j))^{-\eta}\; \frac{h(t,S,D+e_j)}{h(t,S,D)} . \]

This is algebraic (a power, not a transcendental) in \(\pi_j\), but — unlike the log case — it remains coupled to the value ratio \(h(t,S,D+e_j)/h(t,S,D)\). Power utility therefore sits between CARA and log: it removes the transcendental obstruction, so each first-order condition is solvable by low-dimensional root-finding, but the reduced system is nonlinear and does not collapse to a matrix exponential. Log (Kelly) is the \(\eta\to 1\) limit in which the value-ratio coupling vanishes and the system becomes linear.

Findings across the lemmas

The matrix locates each of the three lemmas within each utility family. Linear entries admit an exact (matrix-exponential or Feynman–Kac) solution; algebraic entries are non-transcendental but nonlinear; transcendental entries have no closed form.

Lemma\(\lambda_i\)CARA \(-e^{-\gamma x}\)Log (Kelly) \(\log x\)Power (CRRA) \(\tfrac{x^{1-\eta}}{1-\eta}\)
L1constantnonlinear, transcendental FOC, no closed formlinear, matrix exponential, constant \(\pi^*\)algebraic FOC, nonlinear algebraic system
L2\(a_i + b_i^\top S\)nonlinear + price generator + cross term, no Riccati closurelinear PDE (Feynman–Kac), \(\pi^* = \pi^*(S)\)algebraic FOC on \((S,D)\), nonlinear
L3general \(\lambda_i(S,D)\)nonlinear coupled HJBlinear PDE, state-dependent coefficients, \(\pi^* = \pi^*(S,D)\)nonlinear coupled HJB

The conclusion is clean. Under CARA the multiplicative default makes every lemma nonlinear with a transcendental first-order condition: the plan's deliverable is the precise negative result locating the boundary, and the boundary is L1 itself. Under log utility (Kelly) the same problem is linear throughout — constant wealth-independent fractions, a matrix-exponential solution for L1, and a linear Feynman–Kac equation for L2/L3. Power utility is the tunable middle ground: non-transcendental and closed in low dimension, but nonlinear. The two non-CARA families are outside the plan's stated CARA scope; they are included here because the tractability boundary cannot be stated without them.