# Finite-T Investigation: Why BP Is Estimated Too Strong And FA Too Weak

This note investigates the finite-time regime beyond the tagged local result
`t5-local-tangent-validation`.

## Empirical Decomposition

For larger \(T\), decompose the gap error into BP and FA terms:

\[
\Delta_{\mathrm{gap}}^{\mathrm{err}}
=
\left(L_{\mathrm{FA}}^{\mathrm{pred}}-L_{\mathrm{BP}}^{\mathrm{pred}}\right)
-
\left(L_{\mathrm{FA}}^{\mathrm{emp}}-L_{\mathrm{BP}}^{\mathrm{emp}}\right).
\]

Equivalently,

\[
\Delta_{\mathrm{gap}}^{\mathrm{err}}
=
\underbrace{
\left(L_{\mathrm{FA}}^{\mathrm{pred}}-L_{\mathrm{FA}}^{\mathrm{emp}}\right)
}_{\mathrm{FA\ error}}
-
\underbrace{
\left(L_{\mathrm{BP}}^{\mathrm{pred}}-L_{\mathrm{BP}}^{\mathrm{emp}}\right)
}_{\mathrm{BP\ error}}.
\]

Observed signs:

- fixed-width \(T=50\): BP is estimated too strong; FA is estimated slightly too
  weak;
- width-sweep \(T=1000\): BP is estimated too strong; FA is estimated much too
  weak.

Numerically:

| experiment | BP pred-emp | FA pred-emp | gap pred-emp |
|---|---:|---:|---:|
| fixed-width \(N=128,T=50\) | -0.0233 | +0.0052 | +0.0285 |
| fixed-width \(N=320,T=50\) | -0.0167 | +0.0050 | +0.0217 |
| width-sweep \(T=1000\) | -0.1471 | +0.3652 | +0.5123 |

So:

1. at moderate \(T\), the dominant error is BP fixed-kernel optimism;
2. at long \(T\), FA fixed-kernel pessimism becomes dominant because FA learns
   alignment.

## Related Work

The original NTK theory shows that in the infinite-width limit, gradient
descent is described by a kernel gradient flow and the NTK stays constant during
training. For finite networks the NTK depends on parameters and can change
during training.

Chizat, Oyallon, and Bach's lazy-training view makes the same distinction:
linearized/lazy dynamics are accurate when parameters barely move; outside that
regime, feature learning changes the tangent kernel.

Finite-width studies ask when NTK theory can be trusted. Seleznova and Kutyniok
emphasize that the empirical NTK can change substantially in finite deep
networks, depending on initialization and depth.

The Neural Tangent Hierarchy of Huang and Yau is the closest formal tool for
finite-width kernel evolution: it describes the time derivative of the NTK using
higher-order tangent objects. This is exactly the type of correction needed for
finite \(T\).

For FA/DFA specifically, Refinetti et al.'s "align, then memorise" picture says
that feedback-alignment learning first adapts weights to improve gradient
alignment, then fits data. This explains why fixed \(K_{\mathrm{FA}}(0)\) becomes
too pessimistic at long \(T\).

Surrogate-gradient NTK work is relevant because FA is a surrogate-gradient
learning rule: the forward Jacobian controls output motion while the surrogate
backward Jacobian controls parameter motion.

## Own Derivation

Use squared loss

\[
L(\theta)
=
\frac{1}{2N}\|r(\theta)\|^2,
\qquad
r(\theta)=f_\theta(X)-y.
\]

Let

\[
J_t
=
\frac{\partial f_\theta(X)}{\partial\theta}
\bigg|_{\theta_t}
\]

be the true forward Jacobian, and let

\[
\tilde J_t
\]

be the surrogate Jacobian used by the learning rule. For BP,

\[
\tilde J_t=J_t.
\]

For FA,

\[
\tilde J_t=\tilde J_{\mathrm{FA}}(B,\theta_t).
\]

Full-batch update:

\[
\theta_{t+1}
=
\theta_t
-
\frac{\eta}{N}
\tilde J_t^\top r_t.
\]

Taylor expand outputs:

\[
r_{t+1}
=
r_t
+
J_t(\theta_{t+1}-\theta_t)
+
\frac12
H_{f,t}[
\theta_{t+1}-\theta_t,
\theta_{t+1}-\theta_t
]
+
O(\|\Delta\theta_t\|^3).
\]

Substitute the update:

\[
r_{t+1}
=
\left(
I-\frac{\eta}{N}K_t
\right)r_t
+
\frac{\eta^2}{2N^2}
H_{f,t}
[
\tilde J_t^\top r_t,
\tilde J_t^\top r_t
]
+
O(\eta^3),
\]

where the learning-rule kernel is

\[
K_t=J_t\tilde J_t^\top.
\]

The local \(T=5\) theorem uses

\[
K_t\approx K_0
\]

and ignores the \(O(\eta^2)\) Taylor term:

\[
r_T^{(0)}
=
\prod_{t=0}^{T-1}
\left(I-\frac{\eta}{N}K_0\right)r_0.
\]

The finite-\(T\) correction has two pieces:

\[
\delta r_T
\approx
\sum_{s<T}
A_{T:s+1}
\left[
-\frac{\eta}{N}\Delta K_s r_s^{(0)}
+
\frac{\eta^2}{2N^2}
H_{f,s}[
\tilde J_s^\top r_s^{(0)},\tilde J_s^\top r_s^{(0)}
]
\right],
\]

where

\[
A_{T:s+1}
=
\prod_{u=s+1}^{T-1}
\left(I-\frac{\eta}{N}K_0\right),
\qquad
\Delta K_s=K_s-K_0.
\]

So finite-\(T\) error is not mysterious: it is accumulated kernel drift plus
second-order output curvature.

## Why BP Is Estimated Too Strong

For BP:

\[
K_t^{\mathrm{BP}}=J_tJ_t^\top.
\]

The fixed-kernel prediction assumes the initial tangent features keep their
initial training power. In the experiments, \(L_{\mathrm{BP}}^{\mathrm{pred}}\)
is below \(L_{\mathrm{BP}}^{\mathrm{emp}}\), so the true finite-width BP dynamics
learn more slowly than the frozen \(K_0\) model predicts.

A sufficient explanation is that along the residual direction,

\[
r_t^\top K_t^{\mathrm{BP}} r_t
<
r_t^\top K_0^{\mathrm{BP}} r_t
\]

on average, or that curvature terms partially oppose the first-order output
motion. This is finite-width/lazy-regime breakdown: the initial BP NTK is too
optimistic as a long-horizon proxy.

This matches the literature: finite-width networks can leave the frozen-NTK
regime, and the empirical NTK can change significantly during training.

## Why FA Is Estimated Too Weak

For FA:

\[
K_t^{\mathrm{FA}}
=
J_t\tilde J_t^\top.
\]

At \(t=0\), \(\tilde J_t\) is based on random feedback and is misaligned with
the BP gradient. The fixed-kernel prediction freezes this bad credit-assignment
operator.

But FA training changes the forward weights so that the surrogate direction
becomes more useful. In kernel language, the alignment improves the real part
or residual-direction action of \(K_t^{\mathrm{FA}}\):

\[
r_t^\top K_t^{\mathrm{FA}} r_t
>
r_t^\top K_0^{\mathrm{FA}} r_t
\]

on average after an alignment phase.

Thus \(L_{\mathrm{FA}}^{\mathrm{pred}}>L_{\mathrm{FA}}^{\mathrm{emp}}\): the
frozen random-feedback kernel underestimates FA learning.

This is exactly consistent with "align, then memorise": early training improves
the credit-assignment geometry before memorisation dominates.

## Practical Finite-T Theory Candidate

The next theory should not be a fitted multiplicative scale. The clean finite-T
object is:

\[
r_T^{\mathrm{tv}}
=
\prod_{t=0}^{T-1}
\left(
I-\frac{\eta}{N}K_t
\right)r_0.
\]

This is directly testable by measuring \(K_t\) at checkpoints.

If the product using measured \(K_t\) matches empirical \(T=50\) and \(T=1000\),
then the remaining task is to model \(K_t\). A compressed model could be:

\[
K_t^{\mathrm{FA}}
\approx
K_0^{\mathrm{FA}}
+
\alpha_t
\left(
K_t^{\mathrm{BP}}-K_0^{\mathrm{FA}}
\right)
+
E_t,
\]

where \(\alpha_t\) is not fitted freely but defined from measurable alignment:

\[
\alpha_t
=
\frac{
\langle K_t^{\mathrm{FA}}-K_0^{\mathrm{FA}},
K_t^{\mathrm{BP}}-K_0^{\mathrm{FA}}\rangle_F
}{
\|K_t^{\mathrm{BP}}-K_0^{\mathrm{FA}}\|_F^2
}.
\]

Then one can test whether \(\alpha_t\) itself follows a simple law, for example
an exponential alignment curve:

\[
\alpha_t\approx 1-\exp(-t/\tau_{\mathrm{align}}).
\]

But this is only acceptable if \(\tau_{\mathrm{align}}\) can be predicted from
architecture/data/initialization statistics or independently measured, not
chosen to fit final loss.

## Recommended Diagnostic

For a small set of trajectories:

1. save checkpoints at \(t=\{0,1,2,5,10,20,50,100,\dots\}\);
2. compute \(K_t^{\mathrm{BP}}\), \(K_t^{\mathrm{FA}}\), residuals \(r_t\);
3. measure directional learning rates

\[
\lambda_t^{\mathrm{dir}}
=
\frac{r_t^\top K_t r_t}{\|r_t\|^2};
\]

4. measure kernel overlap

\[
\rho_t
=
\frac{\langle K_t^{\mathrm{FA}},K_t^{\mathrm{BP}}\rangle_F}
{\|K_t^{\mathrm{FA}}\|_F\|K_t^{\mathrm{BP}}\|_F};
\]

5. compare empirical residuals with the measured time-varying kernel product.

This will tell us whether finite-\(T\) is mostly kernel drift or whether output
curvature terms must be modeled explicitly.