Without knowing more about your data and seeing some
regression output, it is difficult to say whether you should drop the linear term, retaining only the quadratic term. However, there is no reason to deprecate this
idea immediately.
If you have predictor variables $x, x^2,$ and $x^3,$ then you may have co-linearity problems. In the following
extremely simple example Minitab output is as follows:
Model Summary
S R-sq R-sq(adj) R-sq(pred)
0.236001 100.00% 100.00% 100.00%
Coefficients
Term Coef SE Coef T-Value P-Value VIF
Constant 12.335 0.510 24.17 0.000
x1 -2.269 0.418 -5.42 0.003 188.59
x2 0.5383 0.0947 5.68 0.002 1016.06
x3 0.36447 0.00625 58.31 0.000 366.71
Regression Equation
y = 12.335 - 2.269 x1 + 0.5383 x2 + 0.36447 x3
A matrix plot of all four variables is as follows:

One may conclude it is best to use only $x^3$ as a
predictor variable:
Regression Analysis: y versus x3
Analysis of Variance
Source DF Adj SS Adj MS F-Value P-Value
Regression 1 83744.0 83744.0 278128.45 0.000
x3 1 83744.0 83744.0 278128.45 0.000
Error 7 2.1 0.3
Total 8 83746.1
Model Summary
S R-sq R-sq(adj) R-sq(pred)
0.548724 100.00% 100.00% 99.99%
Coefficients
Term Coef SE Coef T-Value P-Value VIF
Constant 9.984 0.250 39.90 0.000
x3 0.400237 0.000759 527.38 0.000 1.00
Regression Equation
y = 9.984 + 0.400237 x3
In view of these results, it seems difficult to argue with using $x^3$ alone in a linear regression.
So, this is an indication that there are situations in polynomial regression, in which
it makes sense to drop lower powers of the $x_i.$
(Confession: These are fake data generated according to
$Y_i = 10 + 0.4x_i^3 + e_i,$ where $e_i$ are normal with 0 mean and small variance. The rough idea is that $Y_i$ is the weight of a full cubical container of a certain kind and $x_i$ is the length of one of its sides.)