Dutch eligibility traces

⚠️ Curated Limitations & Negative Results

• The component is distinct from complete true online TD or Sarsa, the general eligibility family and an entire forward-view learning algorithm. In the UAI theorem, forward/backward weight equality also requires its linear weight updates, companion correction and horizon-independent target-difference relation; inserting only this trace into an arbitrary host does not establish that equality.
• The scaled and unscaled conventions must be matched to the host update. Constant-alpha rescaling E=alpha*e does not justify the substitution E_t=alpha_t*e_t for varying step sizes. Division by alpha is undefined at alpha=0. A current zero alpha_t in the scaled recurrence leaves the decayed previous trace when earlier step sizes produced nonzero credit.
• The accumulating/replacing relation is an identity between one-hot update maps evaluated from the same old trace, not a mixture of separately evolved trace histories. It does not extend as a coordinatewise replacing rule to arbitrary overlapping binary or real features. Coincident one-hot trace updates do not make the complete learning algorithms identical.
• Dutch traces are not coordinatewise clipping and need not remain nonnegative or within [0,1]. In the general scaled recurrence, if z=c*phi^T*E_old and h=eta*||phi||^2, then phi^T*E_new=z+h*(1-z). For any starting projection, h in [0,1] keeps the new projection between its old value and one; unrestricted feature norms or step sizes can overshoot or amplify the deviation. The source phrase toward one does not imply exact unit attainment.
• The general UAI coefficient c is a scalar independent of the data horizon, not necessarily a number in [0,1]. Its description as shrinking requires additional magnitude conditions. The transition c*(I-eta*phi*phi^T) has feature-direction factor c*(1-eta*||phi||^2), so the recurrence supplies no unconditional numerical stability, contraction, convergence, variance reduction or performance guarantee.
• The source forward-view derivations concern linear prediction in the learned weights with the specified feature/target sequence and matching initialization. The vector trace formula can be evaluated on supplied features, but generic nonlinear or parameter-dependent feature learning does not automatically inherit those derivations. Dense O(n) bookkeeping does not imply cost proportional only to the current number of nonzero features or a universal wall-clock advantage.
• The 2014 PMLR paper supplies the earlier scaled recurrence; the 2014 UAI paper explicitly names the general form Dutch traces; the 2016 journal article presents unscaled, time-dependent and tabular conventions and attributes the name to van Hasselt and colleagues. This is inspected source lineage, not an exhaustive priority claim. The UAI proof locally omits +x_k in its first F_k restatement although the theorem and later expansion retain those terms, and the 2016 prose after Equation 17 points to Algorithm 2 although the local time-dependent box is Algorithm 4. These bounded display observations are not a certified defect of the trace recurrence, an official erratum or a full proof audit. No author implementation or benchmark reproduction is certified.