1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
|
# Early Predictor Probe
Measured time-varying \(K(t)\) is an oracle diagnostic. To get a predictive
finite-\(T\) theory, we need quantities computable from initialization or the
first few steps.
## Probe Setup
- architecture \(16\to64\to64\to4\);
- \(N=128\);
- target horizon \(T=50\);
- SGD, learning rate \(10^{-3}\);
- 4 init seeds and 8 feedback seeds;
- early probes \(s\in\{1,2,5\}\).
Candidate predictors:
- fixed \(K(0)\) gap prediction;
- FA kernel Frobenius drift:
\[
\|K_s^{FA}-K_0^{FA}\|_F/\|K_0^{FA}\|_F;
\]
- BP kernel Frobenius drift:
\[
\|K_s^{BP}-K_0^{BP}\|_F/\|K_0^{BP}\|_F;
\]
- FA/BP kernel overlap:
\[
\rho_s
=
\frac{\langle K_s^{FA},K_s^{BP}\rangle_F}
{\|K_s^{FA}\|_F\|K_s^{BP}\|_F};
\]
- residual-direction gains:
\[
\lambda_s^{FA}
=
\frac{r_s^\top K_s^{FA}r_s}{\|r_s\|^2},
\qquad
\lambda_s^{BP}
=
\frac{r_s^\top K_s^{BP}r_s}{\|r_s\|^2}.
\]
## Result
Output:
`outputs/early_kernel_predictors_T50_N128_32runs/early_kernel_predictors.csv`
The strongest single predictor of the \(T=50\) empirical gap remains the fixed
\(K(0)\) predicted gap:
\[
\operatorname{corr}
\left(
\Delta L_{50}^{fixedK},
\Delta L_{50}^{emp}
\right)
\approx
0.951.
\]
The fixed prediction is biased high:
\[
\operatorname{MAE}
\left(
\Delta L_{50}^{fixedK},
\Delta L_{50}^{emp}
\right)
=
0.02063.
\]
If we fit only a linear calibration of fixed gap on this single setting, the
in-sample MAE drops to:
\[
0.00517.
\]
Adding an early drift predictor can reduce the in-sample residual further. The
best early predictors were:
- \(s=1\), fixed gap + FA directional drift:
\[
\operatorname{MAE}\approx0.00296;
\]
- \(s=5\), fixed gap + BP Frobenius drift:
\[
\operatorname{MAE}\approx0.00307.
\]
## Interpretation
This probe is useful but not yet a theorem-ready predictor.
The good news:
1. fixed \(K(0)\) already ranks trajectories well at \(T=50\);
2. early drift features explain part of the finite-\(T\) residual;
3. this supports the finite-\(T\) picture:
\[
\Delta L_T
=
F(K_0)
+
\text{early drift correction}
+
\text{higher-order residual}.
\]
The caution:
The linear combinations above are in-sample on one architecture/task setting.
They are not yet valid paper predictors. The next step must test whether a
fixed formula learned or derived from one setting generalizes across \(N\),
width, and feedback seeds.
## Next Direction
The most promising non-oracle predictive object is not raw Frobenius drift. It
is the residual-direction kernel action:
\[
\lambda_t^{rule}
=
\frac{r_t^\top K_t^{rule}r_t}{\|r_t\|^2}.
\]
Reason: loss decrease is controlled by
\[
L_{t+1}-L_t
\approx
-
\frac{\eta}{N}
r_t^\top K_t r_t.
\]
So a finite-\(T\) predictor should approximate the cumulative directional gain:
\[
G_T^{rule}
=
\sum_{t<T}
\frac{\eta}{N}
\lambda_t^{rule}.
\]
For a predictive version, estimate the early slope:
\[
\lambda_t
\approx
\lambda_0+t\dot\lambda_0,
\qquad
\dot\lambda_0
\approx
\frac{\lambda_s-\lambda_0}{s},
\]
then predict
\[
G_T
\approx
\frac{\eta}{N}
\left(
T\lambda_0+\frac{T(T-1)}{2}\dot\lambda_0
\right).
\]
This is more directly tied to loss dynamics than global kernel overlap.
## Cross-N Directional Predictor Check
We also tested a no-fit directional-gain extrapolation across
\[
N\in\{64,96,128,160,192,224,256,320\}
\]
with width fixed at \(64\), target horizon \(T=50\), 2 init seeds, and 4
feedback seeds per init.
The attempted predictor used
\[
\lambda_t\approx\lambda_0+t\dot\lambda_0,
\qquad
\dot\lambda_0\approx\frac{\lambda_s-\lambda_0}{s},
\]
and
\[
L_T
\approx
L_0
\exp\left[
-\frac{2\eta}{N}
\left(
T\lambda_0+\frac{T(T-1)}{2}\dot\lambda_0
\right)
\right].
\]
This failed as a finite-\(T\) gap predictor. Across \(N\), the directional-gap
prediction had low correlation and large negative bias:
| early step | fixed \(K(0)\) corr | fixed MAE | directional corr | directional MAE |
|---:|---:|---:|---:|---:|
| 1 | 0.979 | 0.0241 | 0.124 | 0.2838 |
| 2 | 0.979 | 0.0241 | 0.138 | 0.2964 |
| 5 | 0.979 | 0.0241 | 0.110 | 0.3686 |
So the simple directional exponential formula should not be used.
The reason is likely that a scalar gain
\[
\lambda_t=\frac{r_t^\top K_t r_t}{\|r_t\|^2}
\]
does not preserve residual direction changes. BP and FA rotate the residual in
different ways, and the gap is a difference between two residual trajectories,
not just a difference between two scalar decay rates.
The more viable predictive object is therefore not a scalar exponential loss
model. It is a compressed operator model, for example:
\[
K_t^{FA}
\approx
K_0^{FA}
+
\alpha_t
\left(K_0^{BP}-K_0^{FA}\right),
\]
or a low-rank/residual-subspace version. The scalar order parameter can control
operator interpolation, but the residual update must remain vector-valued:
\[
r_{t+1}
=
\left(I-\frac{\eta}{N}K_t^{compressed}\right)r_t.
\]
This preserves direction changes and is closer to the measured \(K(t)\) product
that worked.
|
