= Solution
The entire term structure is naturally indexed by observation time and maturity. Work under a <risk-neutral measure> $Q$ and let $f(t,T)$ be the <instantaneous forward rate>, with $0\le t\le T$. Define the <short rate> $r_t=f(t,t)$, the <bank account> $B_t=\exp(\int_0^t r_sds)$ and the <zero-coupon bond> price
$$
P(t,T)=\exp\left(-\int_t^T f(t,u)du\right).
$$
An arbitrary Gaussian mean and <covariance> need not make these bond prices consistent. Absence of <arbitrage> requires the discounted traded bonds to be <martingales> under a suitable equivalent measure. We derive the constraint explicitly.
Consider the <Gaussian forward-rate field> $f(t,T)=m(t,T)+X(t,T)$ with deterministic mean and centered <Gaussian random field> satisfying
$$
\operatorname{Cov}(X(s,T),X(t,U))=c(s\wedge t;T,U),\qquad c(0;T,U)=0.
$$
The maturity kernel is symmetric, and its increments in the first parameter must be positive semidefinite kernels. Assume continuity, separability and sufficient integrability for the following mean-square integrals and maturity derivatives. This <covariance> gives independent observation-time increments, while permitting correlated fluctuations across all maturities. The <filtration> contains the history of the whole observed forward curve.
For a fixed maturity $T$, collect the past <short rate> and the present curve into the <integrated Gaussian forward-rate process>
$$
Y(t,T)=\int_0^T X(t\wedge u,u)du,
\quad A(t,T)=\int_0^t m(u,u)du+\int_t^T m(t,u)du.
$$
Then $P(t,T)/B_t=e^{-A(t,T)-Y(t,T)}$. Its centered Gaussian <variance> is
$$
v(t,T)=\int_0^T\int_0^T c(t\wedge u\wedge w;u,w)du\,dw.
$$
For $s<t$ and an observed field coordinate $X(z,U)$ with $z\le s$, the <covariance> of $Y(t,T)-Y(s,T)$ with it is zero: the integrand is
$$
c(t\wedge u\wedge z;u,U)-c(s\wedge u\wedge z;u,U)=0.
$$
The <uncorrelated jointly Gaussian variables are independent> principle therefore makes the increment independent of the earlier field <filtration>. Its <variance> is $v(t,T)-v(s,T)$, since the same <covariance> calculation gives $\operatorname{Cov}(Y(t,T),Y(s,T))=v(s,T)$. Conditional Gaussian exponential <expectation> now gives
$$
\mathbb E_Q[e^{-A(t,T)-Y(t,T)}\mid\mathcal F_s]
=e^{-A(t,T)-Y(s,T)+[v(t,T)-v(s,T)]/2}.
$$
Hence the exact bond-martingale condition is
$$
A(t,T)-A(0,T)=\frac12v(t,T).
$$
Differentiating in maturity, using symmetry of the <covariance>, gives the <Gaussian forward-rate covariance drift restriction>
$$
\boxed{m(t,T)=f(0,T)+\int_0^T c(t\wedge u;u,T)du.}
$$
This is necessary and sufficient. For sufficiency, the derivatives of the half-variance identity agree by the displayed restriction. Its integration constant also agrees at $T=t$, because
$$
A(t,t)-A(0,t)=\int_0^t\int_0^u c(w;w,u)dw\,du=\frac12v(t,t).
$$
Thus the full identity follows for every $T\ge t$, and the <conditional expectation> calculation proves a true <martingale>, not merely zero formal drift. The initial mean $f(0,T)=-\partial_T\log P(0,T)$ fits the observed initial curve exactly; thereafter the mean adjustment is fixed by the <covariance>.
If $c(t;T,U)=\int_0^t k_s(T,U)ds$ is time-differentiable, the forward drift is $a(t,T)=\int_t^T k_t(T,u)du$. Writing
$$
k_t(T,U)=\langle\sigma(t,T),\sigma(t,U)\rangle
$$
with deterministic finite- or Hilbert-space factor loadings gives the <Heath-Jarrow-Morton model>
$$
df(t,T)=\left\langle\sigma(t,T),\int_t^T\sigma(t,u)du\right\rangle dt
+\langle\sigma(t,T),dW_t^Q\rangle.
$$
With $\Sigma(t,T)=\int_t^T\sigma(t,u)du$, <Itô formula> independently verifies bond drift $r_t-\int_t^Ta(t,u)du+\|\Sigma(t,T)\|^2/2=r_t$. Finite-factor models give low-rank maturity correlations; a genuine random field or an infinite factor space permits a much richer <covariance> structure. Gaussian forward rates and <short rates> may be negative, although the exponential bond prices stay positive. Under a physical measure, a <Girsanov theorem> market-risk-premium change adds $\langle\sigma,\lambda\rangle$ to the forward drift; deterministic premia preserve Gaussianity, whereas arbitrary adapted premia need not.
For a concrete two-parameter example, let $X$ be a standard <Brownian sheet>, so $c(t;T,U)=t\min(T,U)$. The restriction gives
$$
m(t,T)=f(0,T)+\frac12tT^2-\frac16t^3\qquad(t\le T).
$$
Indeed $\int_0^T\min(t,u)u\,du=t^3/3+t(T^2-t^2)/2$. This demonstrates directly how a <Gaussian random field> determines the compensating forward drift.
For the one-factor constant-volatility case, $\sigma(t,T)=\eta$, the <Constant-coefficient Ho-Lee model> has
$$
f(t,T)=f(0,T)+\eta^2(tT-t^2/2)+\eta W_t^Q,
\qquad r_t=f(0,t)+\eta^2t^2/2+\eta W_t^Q,
$$
and
$$
P(t,T)=\frac{P(0,T)}{P(0,t)}
\exp\left[-\eta(T-t)W_t^Q-\frac12\eta^2tT(T-t)\right].
$$
These follow by integrating the explicit curve, not by assigning the short-rate drift independently. Exponentially decaying loadings $\eta e^{-\kappa(T-t)}$ instead give a Gaussian mean-reverting <Hull-White model>, with bond-rate loading $(1-e^{-\kappa(T-t)})/\kappa$.
A practical advantage is an explicit <Gaussian bond-option formula>. For a call of strike $K>0$ exercised at $T$ on a bond maturing at $U>T$, use the <T-forward measure> with numéraire $P(t,T)$. The forward bond ratio $F_t=P(t,U)/P(t,T)$ has a <lognormal distribution> under that measure, with remaining integrated <variance>
$$
q=\int_t^T\|\Sigma(s,U)-\Sigma(s,T)\|^2ds.
$$
Its <martingale> property makes the conditional log mean $\log F_t-q/2$. Evaluating the positive-part lognormal integral gives
$$
\boxed{C_t=P(t,U)\Phi(d_1)-KP(t,T)\Phi(d_2),\quad
d_1=\frac{\log(F_t/K)+q/2}{\sqrt q},\quad d_2=d_1-\sqrt q.}
$$
For $q=0$, the limit is $(P(t,U)-KP(t,T))^+$. The forward measure accounts for the stochastic <short rate>; discounting by one deterministic interest rate would generally be wrong. Calibration selects a positive-semidefinite <covariance> structure fitting observed rate co-movements and option prices, with the no-arbitrage mean restriction imposed afterward. With more independent factors than the available traded bond exposures span, the market is incomplete and additional claims need a specified pricing measure or risk-premium model. The Gaussian field is therefore a flexible curve model with explicit consistency and pricing equations, rather than a claim that all possible rate risks are hedgeable.
Back to article page
